AI automation agency for ops teams — what 'custom' actually costs

The AI automation agency SERP is split into two kinds of pages, and neither of them is written for the buyer. The top third are "$0 to $30k/month in 90 days" programs teaching people how to start an agency. The bottom third are generic listicles sorted by who paid for backlinks. The middle third — studios that actually ship production AI agents to ops teams — is where the intent mismatch gets expensive, because a course-graduate funnel shop and a code-first studio cost roughly the same in the proposal and diverge 10x in outcomes.

We're Autoolize, an AI automation studio and a member of the Anthropic Claude Partner Network³. Across 40 production AI agents shipped for ops teams between 2024 and 2026 — invoice OCR, inbound support triage, vendor reconciliation, lead enrichment, research agents — we've seen every archetype in the market from the other side of the proposal. This is the buyer's guide we wish existed when new clients come to us after a prior agency engagement that didn't stick.

Two numbers anchor the post. First, the MIT NANDA GenAI Divide: State of AI in Business 2025 report found that roughly 95% of enterprise generative AI pilots fail to produce measurable P&L impact, despite $30–40 billion in aggregate investment¹. Second, McKinsey's State of AI in 2025 survey shows 88% of companies now use AI regularly in at least one function, but only 39% report enterprise-level EBIT impact from AI — and in most cases that impact accounts for less than 5% of total EBIT². Translation: the market is flush with capital and empty of outcomes. The agency you pick is the single decision that tips your project into the minority of AI work that ships measurable returns, or leaves it with the majority that doesn't.

If you're comparison-shopping, skim §1 for the six archetypes at a glance and §5 for the full rundown. If you want honest pricing before you scope, book a strategy call — we'll share two redacted SOWs from recent ops deployments so you know what a real build invoice looks like.

Quick overview — 6 agency archetypes at a glance

The AI automation agency category has six recognizable shapes in 2026. Same SEO page title, radically different businesses. We sort them by how well each fits an ops buyer — the first two are the ones to shortlist, the next two can work in narrow cases, the last two are the ones a buyer should walk away from unless the price is artificially low.

#	Archetype	Typical cost	Best for	Red flag
1	Code-first AI studio	$20k–$120k per agent	Production ops agents with SLAs	No public case studies with dollar impact
2	Platform vendor's services arm	$30k–$250k per project	Teams already committed to the vendor	Zero pretense of vendor neutrality
3	Generalist dev shop reskin	$50k–$200k per project	Custom UI wrapping an AI component	"AI automation" not on their site pre-2024
4	Productized funnel shop	$8k–$40k + retainer	Short-run Make/n8n automations	Same landing page as 50 others, no named engineer
5	Course-to-agency graduate	$3k–$15k/month retainer	Almost no ops-production use case	Founder posts more about "starting an agency" than about the work
6	White-label reseller	Whatever they quote you	Never	The engineer on the demo call won't be on your project

The shortlist. For a production ops agent — invoice OCR, inbound triage, vendor reconciliation, lead enrichment, internal research — you want archetype 1 (code-first studio) or archetype 2 (platform vendor's services arm, only if you've already picked the platform). Everything else is either too thin or not actually a fit for ops production work, however polished the landing page.

The litmus test. Before reading §5 in full, run three fast checks on any agency you're considering. (1) Ask for a production case study with a dollar impact figure and a measurement window. If they can't produce one inside a day, archetype 4 or 5. (2) Ask for the stack by name — model, agent framework, observability. If the answer is only "Make" or "n8n," archetype 4 or 5. (3) Ask who specifically will be writing the code. If the engineer on the demo call isn't the same person on your project, archetype 6.

Three yeses gets you onto the shortlist. Anything less is a pass, regardless of price.

How to tell a studio from a course

The AI automation agency SERP is the most intent-mismatched head term in the AI services space. Roughly half the top-10 results are selling courses to aspiring agency owners — programs with titles like "AI Automation Agency Blueprint" and "$0 to $30k/month in 90 days." The other half are listicles of agencies, many of which are themselves businesses started inside one of those same programs. The search intent is split roughly 60-40 between "how do I start an agency" and "which agency do I hire," but the pages don't label themselves clearly, so the buyer has to do the sorting.

The fastest filter is who the website is actually selling to.

Selling to buyers. Site copy names ops outcomes (cost savings, cycle time, SLA uplift), not revenue numbers for the agency. Case studies name real clients, dollar impact, and the measurement window. The founder's author page lists technical writing — agent architecture, eval harness design, production failure modes — not LinkedIn motivational posts. Pricing is either published or quoted within 48 hours of discovery. The proposal names specific tools: Claude Agent SDK, OpenAI Agent Builder, LangGraph, Anthropic Skills, specific eval frameworks.

Selling to aspiring agency owners. Site copy names agency revenue ("scale your agency to $30k/month"), income proof, community access, and "done-for-you templates." The hero CTA is "Join the program" or "Book a discovery call" framed around the buyer's own revenue goal. Case studies describe the agency's cohort, not ops clients. The founder's most-recent content is pitching the program itself. Tools named are almost always low-code first — Make, n8n, Zapier, Airtable — because the pitch is "you don't need to code." Agency-building SOPs, cold-outbound scripts, and sales playbooks are the upsells.

Both businesses are legitimate and profitable. Only one of them is the business you want handling your ops agent. The other is a cohort business that happens to sometimes take on agency work so the program has case studies to sell.

The nastier version of this: the alumni. The agencies a buyer is most likely to mistake for an archetype-1 studio are archetype-5 course graduates — operators one or two years out of a program, with a clean website, 2–5 client logos, and a polished Loom demo. Those agencies are often real businesses; the question is whether the work they ship matches an ops production bar. Three signals separate the strong alumni from the weak ones. First, they publish technical content about the work, not motivational content about the journey. Second, they quote fixed-price per-agent, not month-to-month retainers that never end. Third, they've kept at least one client past the 12-month mark — ask for a reference where the engagement has renewed at least once.

The course-graduate business model has a quiet problem for ops buyers: a 90-day program can teach Make, n8n, and prompt design, but it cannot teach the parts that matter at production scale — eval-gated deployment, tool-use retry strategy, cost and latency budgeting under load, observability instrumented for SLAs, or the agent-SDK primitives (subagents, Skills, canvas nodes) that separate prototypes from production systems. Those parts are learned by shipping, measuring, and iterating on real production traffic, which takes years. A strong studio can be two years old; it cannot be two months old.

One more filter: case-study quality. A strong case study reads like an engineering postmortem. It names the workflow, the baseline (time, cost, error rate), the agent design, the eval harness, the rollout, the failures, and the post-launch metric at 30, 60, and 90 days. A weak case study reads like an ad — soft adjectives, no numbers, no named systems, a stock photo of a smiling person. If the case studies don't sound like postmortems, the work the agency ships doesn't sound like production.

A note on "featured in Forbes" and similar. Press badges on agency websites are almost always paid placement in 2026, not editorial coverage. Treat them as marketing spend, not credibility signals. Ask for the original article URL; if it's a paid-contributor column or a "business spotlight" section that charges for inclusion, it tells you the agency has a budget for PR, not that the agency is good at shipping agents. The credibility signals that actually matter are Claude Partner Network or OpenAI partner status³, named engineering leadership with public technical writing, and production case studies with dollar impact.

What "custom AI agent" actually means in 2026

The phrase "custom AI agent" is so overloaded it has lost most of its signal. Pin it to something concrete. A production AI agent in 2026 is a software system with five parts. A custom agent has all five parts, and at least two of them are built specifically for your workflow rather than pulled from a template.

Part 1 — The model layer. The frontier model (Claude Sonnet 4.6, Claude Opus 4.7, GPT-5.4, Gemini 2.5) plus a router layer that picks between a small cheap model (Claude Haiku 4.5, GPT-5-mini) and a full specialist per request. Custom here means tuned prompting — a system prompt tailored to your data and tone, not a generic template — and a router threshold calibrated to your cost/quality budget.

Part 2 — The agent framework. The code runtime — Claude Agent SDK, OpenAI Agents SDK, LangGraph, CrewAI, or a thin custom wrapper. This is where subagents, tool definitions, retry logic, streaming, and state management live. Custom here means the agent is structured in a way that reflects how your ops team thinks about the workflow, with explicit subagent boundaries (we wrote about this in our Claude Agent SDK production playbook) rather than one monolithic prompt doing everything.

Part 3 — The tools. Every external action the agent takes — querying your CRM, writing to NetSuite, searching a vector index, reading a PDF, posting to Slack. Custom here means tools that match your actual system surface, not wrapped SaaS connectors that approximate it. If your accounting system is old-enough-to-matter NetSuite with three decades of custom fields, a generic NetSuite connector won't cut it; a custom tool that reads the exact saved searches your ops team uses will.

Part 4 — The data layer. Everything from input preprocessing (OCR, entity extraction, normalization) to retrieval (vector embeddings, hybrid search, reranking) to write paths (idempotency keys, audit logs, approval queues). Custom here means indexes tuned on your corpus, extraction schemas that match your document types, and write paths that match your compliance requirements.

Part 5 — The eval + observability harness. The production infrastructure that catches regressions: a suite of golden traces the agent must still pass, real-time drift detection, per-request cost and latency logging, alerts tied to your SLAs. Custom here means eval cases written from your actual failure modes, dashboards aligned to your SLAs, and runbooks your on-call engineer can follow at 3 a.m.

An agent is custom when at least two of those five parts — usually the tools and the eval harness — are built for your workflow rather than reused from a template. That's a meaningful build, usually 3–6 weeks of engineering for a single workflow.

Here's what "custom" is not. A Make or n8n scenario that calls OpenAI once per trigger is not a custom AI agent; it's a templated automation with an LLM node. A ChatGPT custom GPT with a system prompt and three actions is not a custom AI agent; it's a custom prompt. A Zapier Agent with natural-language workflow instructions is not a custom AI agent; it's a customer-configurable automation. None of these are bad — they're cheap, quick, and right for a lot of workflows. They're just not what a buyer expects when the proposal says "custom AI agent" and the price is $60k.

Why the distinction matters commercially. The three categories (templated automation, customer-configured agent, custom agent) span a price range of roughly 50x. If you pay custom-agent pricing for a templated automation, you overpay by 30x. If you buy a templated automation when your workflow actually needs a custom agent, the automation will not handle the edge cases that matter to ops, and you'll rebuild it as a real agent in 6–12 months. Both failure modes are common. Both cost much more than a careful scoping conversation.

Three fast reality checks tell you which category your workflow falls into. First, edge case density. If the workflow's edge cases are more than ~10% of volume and they're where the cost actually sits, you need a custom agent — templates can't reason about edge cases they haven't been templated for. Second, system surface. If the agent needs to touch more than three internal systems with custom schemas, a templated automation will collapse under the integration weight. Third, SLA sensitivity. If a single bad output has a dollar cost over ~$500 (a misrouted invoice, a misclassified support ticket during an outage, a wrong bill to a customer), the eval harness you get only with a custom build is the thing that pays back the price difference.

If none of those three apply, buy the templated automation. If one or more does, hire an archetype-1 studio.

The 3 pricing models and which works for ops

Agencies price AI work three ways in 2026. Each has a reasonable use case and a failure mode. The smart move is matching the pricing model to the shape of your workflow, not to whichever one is in front of you.

Model 1 — Fixed-price per agent. You agree on scope, deliverables, and a single number. Typical range is $15k–$120k per agent, with the low end for a single narrow workflow (inbound triage on 1,000 tickets/week) and the high end for multi-system agents with compliance requirements (invoice OCR writing to NetSuite with audit logs and reviewer queues). Payment usually splits 30/40/30 across discovery, MVP, and production rollout.

When it fits. The workflow is scoped and measurable — you can write the outcome in one sentence, name the systems, and name the metric. This is the cleanest model for ops buyers because the studio has to commit; if they blow the scope, they eat the overrun. Our own default.

When it doesn't. The workflow is exploratory — you're not sure what the agent should even do. Fixed-price forces a scope too early and you end up paying for the wrong thing. Either scope a discovery engagement first ($1,500–$15k) or use Model 3.

Red flag. The fixed price doesn't include an eval harness or a hypercare window. That means the studio is optimizing for "ship and move on," and the production failure rate will land on you — usually in the first 30 days after handoff.

Model 2 — Monthly retainer. You pay a flat monthly fee — usually $4k–$25k/month — and the studio treats your workflow as an ongoing engagement. Typical shape is 3–6 months for the build, then a smaller maintenance retainer ($1.5k–$5k/month) for iteration, drift-fixing, and new-workflow scoping.

When it fits. You have a pipeline of 3+ workflows, not a single one, and you want the studio to build, hand off, stay on for iteration, then take the next workflow. The long engagement lets the studio learn your data and systems once and reuse that knowledge across workflows — usually saves 20–30% on the second and third agents versus a cold fixed-price engagement.

When it doesn't. You only have one workflow. Paying a retainer for one agent is slower and more expensive than a fixed-price engagement. It also creates a weird incentive where the studio doesn't want the engagement to end.

Red flag. The retainer has no exit clause, no deliverable schedule, and no "what we built this month" report. That's a billable-hours business wearing the costume of a retainer.

Model 3 — Productized audit (audit-to-build). A fixed-price upfront audit — $1,500–$7,500, delivered in one to two weeks — that produces a scoping document, architecture recommendation, and a fixed-price proposal for the build if you want to proceed. Some studios (including us) refund the audit fee against the build if you sign.

When it fits. You're early in evaluating whether AI agents are even the right answer for your workflow, and you want an expert diagnosis before committing $60k. A strong audit saves 2–4 weeks of wasted scoping on the build side if you do move forward, and saves the full build cost if the honest answer is "this doesn't need an agent, it needs a Zapier scenario."

When it doesn't. You've already done the diagnosis internally, you have a clear workflow spec, and you're ready to buy a build. The audit step adds 1–2 weeks without changing the outcome. Ask for a fixed-price engagement directly.

Red flag. The audit price is more than 10% of the likely build price, or the audit conclusion is always "yes, you need the build." A productized audit is legitimate when the studio has a non-zero rate of recommending against a build on audit calls.

Autoolize's model. We default to Model 3 into Model 1. A $1,500 one-week audit (refunded against any build contract over $20k, or you walk with the scoping doc and we're done), followed by a fixed-price build engagement with an eval harness and a 30-day hypercare window included in the quote. We run a Model 2 retainer for a handful of long-standing clients where we've already shipped three or more agents and they want rolling capacity — it isn't our default because the incentives drift when the meter runs indefinitely.

The right pricing model for an ops buyer. In order of fit: Model 3 (audit) if you're uncertain about scope, then Model 1 (fixed-price) for each agent, and Model 2 (retainer) only if you have an established relationship and a pipeline of work. Reverse that order — retainer first, no audit, no fixed-price — and you're almost certainly looking at archetype 4 or 5.

6 AI automation agency archetypes, compared

Six shapes. Same category on Google. Radically different businesses. We order them by how well each matches an ops production agent — the first two are the ones to shortlist, the next two can work in narrow cases, the last two are mostly landmines. For every archetype we name the scope, the pricing pattern, the proof-points to check, and how the archetype compares to the code-first studio model (archetype 1) that's the default fit for ops.

Archetype 1 — Code-first AI studio

Scope. Builds and ships production AI agents using code-first agent SDKs — Claude Agent SDK, OpenAI Agents SDK, LangGraph — with custom tool definitions, eval harnesses, and observability. Typical engagements are single-workflow agents (invoice OCR, inbound triage, vendor reconciliation) or small portfolios of 3–6 agents for one client. The work is done by engineers who've shipped LLM production systems before, not by freshly-rebranded generalists.

Pricing pattern. Fixed-price per agent, $20k–$120k. Productized audit as the entry point ($1,500–$7,500, often refundable). Retainers only for long-standing relationships with 3+ shipped agents.

Proof-points to check.

• Named production case studies with dollar impact and measurement window (minimum two, ideally four).
• Public writing about agent architecture, eval harness design, and production failure modes — technical, not motivational.
• Claude Partner Network membership³, OpenAI partner status, or equivalent vendor accreditation that requires technical verification.
• Engineers named on the website, with public GitHub or published technical work.
• References from clients past the 12-month mark.

Compared to the studio model. This is the studio model. Everything below is a variation or a departure.

Red flags. No public case study with numbers. Generic "AI automation" landing page without a technical write-up anywhere. Founder's content is motivational or agency-building rather than technical.

Examples. Autoolize (yes, we're biased — but we're an archetype-1 shop and we're writing this for buyers who want the archetype, not the name). Other Claude Partner Network members, OpenAI partner shops, and small engineering-led boutiques that publish production postmortems on their blog.

Archetype 2 — Platform vendor's services arm

Scope. Services organization inside a platform vendor — Sierra's professional services, Lindy's implementation team, Decagon's deploy-led engineering, Anthropic's or OpenAI's own solutions architects on large accounts. Builds custom deployments on the vendor's own stack only. Typical engagement is an enterprise rollout tied to a seat-license or volume commitment with the vendor.

Pricing pattern. $30k–$250k+, often bundled with the platform license. Fixed-price for initial rollout, then tied to platform subscription renewal. Heavily account-manager-mediated.

Proof-points to check.

• The vendor is a real production-grade platform with a meaningful share of enterprise deployments, not a seed-stage company doing services to survive.
• The services team has engineers, not only solution architects — there's someone who can write code on your project, not only someone who maps requirements to the platform's existing features.
• Public case studies with named customers and outcomes on the vendor's own stack.

Compared to the studio model. Strong fit when you've already picked the platform and committed to its ecosystem. Weak fit when you haven't — vendor services arms are not going to recommend a competitor's stack even when it's the right answer. You also lose flexibility: you're locked into the vendor's update cadence, their pricing changes, and their platform limits. Typical TCO is 20–40% higher than an archetype-1 studio building on a generalist agent SDK, but you get vendor escalation channels in exchange.

Red flags. The vendor is pre-enterprise (earlier than Series B or equivalent), the services org is brand new, or the engagement requires a multi-year license commitment before the build starts. All three together means you're paying to de-risk the vendor's business, not yours.

Examples. Sierra deployments for enterprise CX teams, Lindy's implementation team for SMB automation, Decagon's deploy-led team for support agents. All real businesses; all the right choice only if the platform itself is the right choice for your workflow.

Archetype 3 — Generalist dev shop reskin

Scope. Software dev shops (10–200 engineers, often Eastern-Europe-, India-, or Latin-America-based) that built custom web and mobile apps through the 2010s, added "AI automation" to their services page in 2023–2024, and now run LLM projects alongside their other work. Strong at custom UI and infrastructure; mixed at agent-specific disciplines (eval harness, subagent design, tool-use retry strategy, cost and latency budgeting).

Pricing pattern. Time-and-materials or fixed-price at $50–$200/hour blended rates, $50k–$200k per project. Longer engagements than archetype 1 because the dev shop model bills by hours, not by shipped agent.

Proof-points to check.

• "AI automation" content dated before 2023 on the website, with archived case studies. If every AI content piece is post-GPT-4 and reads like it was written last quarter, the skill is freshly bolted on.
• Engineers with AI-specific writing or speaking history, not just generic senior backend engineers marketed as "AI engineers" on the team page.
• Case studies that go past chatbot-on-a-website into production agent work with measurable outcomes.

Compared to the studio model. Strong fit when the project is mostly custom app infrastructure with an AI component — say, a custom internal tool with an embedded copilot that calls an LLM through a well-defined interface. Weak fit when the project is an agent-first workflow (invoice OCR, support triage) where eval-gated deployment and tool-use reliability are the whole game. The dev shop will ship the project; it will just ship it with the wrong architecture for an ops agent, and you'll pay for the education.

Red flags. The engineer on the discovery call is pitched as an "AI engineer" but their public portfolio is Node.js backends and React apps. "AI automation" was added to the services page in the last 18 months. Proposed stack is ambiguous — "we'll use GPT-4 / Claude / whatever works best" — rather than opinionated.

Examples. Large generalist shops with a recent AI practice — the kind of firm that also does mobile apps, custom CMS work, and staff augmentation. Real businesses, real engineers, just not archetype-1 fit for an ops-first agent.

Archetype 4 — Productized funnel shop

Scope. One-to-three-person shops selling "Done-for-you AI automation" as a productized offer — scoped to a narrow use case (lead enrichment, inbound DM automation, content repurposing), usually built on Make, n8n, Zapier, or Airtable plus OpenAI. Heavy cold outbound, heavy founder LinkedIn presence, same landing page template as dozens of other shops. Ships fast because the offer is templated; the template is what the client is actually buying.

Pricing pattern. $8k–$40k setup + $1k–$5k/month retainer. Payment often demanded fully upfront. Retainer has no deliverable schedule — effectively a maintenance fee. Occasionally "revenue share" models on sales-adjacent workflows.

Proof-points to check.

• Video demos of actual workflows running on named clients, not just generic templates.
• Founder's LinkedIn content ratio: how much is technical vs. how much is "I hit $20k MRR, here's how I did it."
• Client logos that can be independently verified (not only "Top 10 brand" with a logo nobody has seen on the brand's actual site).

Compared to the studio model. Strong fit for genuinely light workflows — a four-step Make scenario that drops a form fill into Clearbit then a CRM. Not a fit for anything that needs eval-gated deployment, tool-use retries, or subagent decomposition. The cost gap between a funnel shop and an archetype-1 studio looks big at proposal time (funnel shop is cheaper), collapses by year one (you'll rebuild the Make scenario as a real agent), and inverts by year two (you'll have paid for both).

Red flags. Same landing page design as 20 other shops you've seen. The founder's most-retweeted content is about growing the agency. Proposal is heavy on "we'll build you a custom AI system" but the actual stack is low-code. Payment demanded fully upfront. Retainer is mandatory even though the original deliverable should be finished.

Examples. The agencies sold into existence by "AI automation agency" programs in 2024–2026. Hundreds of them. Real businesses with real workflows for the right small-scope use case — not an ops-production fit.

Archetype 5 — Course-to-agency graduate

Scope. A single operator (or founder plus one contractor) inside their first 12–24 months post-program. Website polished from a template, 2–5 client logos, one or two Loom demos. Works mostly on productized scopes similar to archetype 4, but individual — not templated across a team. Quality varies wildly based on the operator's prior background; a former software engineer who took the program can be decent, a former marketer running the same program can be painful.

Pricing pattern. $3k–$15k/month retainer, monthly, no exit clause. Occasionally a $5k–$20k setup fee. Hourly rates when asked directly usually come in around $100–$200/hour, which tells you the shop is priced for agency owners as peers, not for enterprise buyers.

Proof-points to check.

• Founder's background before the program — software engineering, ops, or unrelated.
• Whether any case study references a production deployment past the 6-month mark (most don't).
• Whether the stack list includes any code-first agent framework, or only low-code (Make, n8n, Zapier, Airtable).

Compared to the studio model. Typically not a fit for ops production work. Can be a fit for a short-scope internal-facing automation where the cost is low enough that a second rebuild in 12 months is acceptable, and where no production SLA depends on the agent. Ask yourself: if this agent breaks silently for a week, what does it cost us? If the answer is "negligible," archetype 5 can work. If it's "more than $10k," shortlist an archetype 1 instead.

Red flags. Founder's most-recent content is about "building an AI agency" rather than "building an AI agent." Website features a strong "community" or "program alumni" badge. The engineer on the demo call is the founder and will also be the engineer on your project — which is honest but limits bandwidth severely. The proposal names Make or n8n as the architecture and describes them as "enterprise-grade."

Examples. Not going to name them, because the specific names churn quarterly as new cohorts graduate. The archetype is the point, not the brand.

Archetype 6 — White-label reseller

Scope. Sells itself as a premium AI automation agency; the actual build is outsourced to a second agency (often archetype 3 or 4) or a marketplace. The reseller does sales, account management, and sometimes a thin discovery layer; the engineering is blind-subcontracted. The engineer on your demo call is not the engineer on your project — usually they never will be, and you never meet them.

Pricing pattern. Opaque. Quote comes in at whatever the market bears — sometimes undercut to undermine archetype-1 proposals, sometimes marked up above archetype 1 on promises of "enterprise process."

Proof-points to check.

• Ask to meet the engineer who will own the work, before signing. A white-label reseller will dodge or substitute. An archetype-1 studio will book the call.
• Ask for the underlying partner's name. If the answer is "our engineering team is proprietary," you're buying from archetype 6.
• Cross-check case studies against public records — who actually shipped the work, and who got the testimonial.

Compared to the studio model. Not a fit for any workflow where you need direct engineering communication or post-launch iteration on production issues. The added layer costs money and adds failure modes without adding any engineering capability the direct studio couldn't provide.

Red flags. The engineer on the demo call won't be on your project. Team page shows founders and account managers but no engineers. "Our enterprise process" is pitched heavily, but the output is indistinguishable from archetype-4 work. Proposal writing reads as templated sales copy rather than a technical specification.

Examples. Mostly intentionally hidden. You'll find them by asking the filtering questions above and watching the answers drift.

AI automations vs bespoke builds — when service beats SaaS

The second-most-common search that lands on a page like this is ai automations — plural, generic, and high-volume (around 9,900 searches/month, per our keyword research). Buyers typing that query are often not sure whether they need an agency at all. A lot of them need SaaS. This section is the decision tree.

Buy a SaaS AI product when three conditions hold. First, your workflow touches at most three internal systems, and they all have published APIs or first-party connectors on the SaaS's integration list. Second, the edge cases you care about are inside the SaaS's pre-built feature set — you're using the vendor's templates, not asking them to build something new. Third, the cost of a bad output is low enough that you don't need a custom eval harness — you're fine with the vendor's quality floor and their standard SLA.

Classic fits: Zapier for simple workflow automations (trigger → action → done), Make for multi-step scenarios with branching, Zendesk Answer Bot or Intercom Fin for first-line support deflection on standard-shape tickets. These are fine choices, and a good agency will sometimes recommend them over a custom build — that's actually the quickest signal of an honest agency.

Build a custom agent (hire an agency) when one of three conditions holds. First, the workflow touches 4+ internal systems, some of them legacy or custom-schema, and the SaaS integration library doesn't cover the set. Second, the edge cases are more than ~10% of volume and they're where the dollar impact lives — a SaaS template won't reason about edge cases it wasn't templated for. Third, the cost of a bad output is high enough (usually >$500 per bad output) that a custom eval harness is the thing that actually pays back.

The cost math isn't subtle. A SaaS plan is $50–$2,000/month. An archetype-1 agency build is $20k–$120k upfront plus ~$500–$2,500/month in ongoing model API and infra costs. Break-even depends entirely on whether the SaaS handles your edge cases. If it does, SaaS is 10–50x cheaper and you should buy it. If it doesn't, the SaaS's failed-workflow cost will overrun the custom build's one-time fee inside 3–6 months.

The common failure: buying SaaS when you need bespoke. A 90-person company runs 650 invoices/day through a generic OCR SaaS. The SaaS handles 85% cleanly and dumps the other 15% into a manual-review queue. A finance analyst spends ~12 hours/week cleaning up the remaining 98 invoices/day. Loaded cost of that analyst time: roughly $2,500/week, or $130k/year. The SaaS bill is $12k/year. Net cost is $142k/year; the SaaS label on the dashboard says "we saved you 85% of the work." A custom invoice OCR agent that handles 96%+ first-pass (where our own deployments land) cuts the queue to ~20 invoices/day and ~2.5 hours/week of analyst time — net cost drops to under $25k/year. Payback on a $60k custom build lands inside 8 months.

The reverse failure: hiring bespoke when SaaS would've worked. A 30-person agency commissions a $45k custom lead enrichment agent when Clearbit + Zapier + a 15-line GPT function would've handled the workflow with equivalent quality. The custom agent works fine, but the buyer overpaid by 40x and now owns a codebase they don't have the engineering team to maintain. This failure is less dramatic day-to-day but is the one we see more often on intake calls — ops buyers convinced by a course-graduate proposal that a custom agent is the answer when a thoughtful SaaS pick would've been.

Our rule for the first 30 minutes of a scoping call. We ask three questions before we talk about a build. (1) What's the volume and what's the edge-case percentage? (2) What are the systems the workflow touches? (3) What's the cost of a bad output? If the answers point to SaaS, we say so and end the call — we'd rather you spend $12k/year with Zapier than $60k with us on the wrong architecture. That's the cheapest free consulting we can give, and it's also the signal you want from any studio before you sign.

What Autoolize does differently

We're an archetype-1 studio. The differentiators below are what separates our engagements from the default archetype-1 pitch, not what separates us from archetypes 4–6 (that's the rest of the post).

Anthropic Claude Partner Network membership. We're a member of the Claude Partner Network, the Anthropic partner program launched March 12, 2026 with a $100M commitment to partner training, engineering support, co-marketing, and the Claude Certified Architect credential track³. Membership alone is free and opt-in — it's not a quality bar by itself, so treat "we're a partner" as a baseline, not proof. The signals inside the program that actually matter for buyers are two: first, partners get a dedicated engineering escalation path with Anthropic, which cuts the clock when something breaks in production — a Claude outage or a new-model regression doesn't queue behind generic support. Second, the Claude Certified Architect exam is a technical credential held by individual engineers (not firms). Any agency that ships seriously on Claude should have at least one certified engineer; asking for that credential is the cleanest vendor-verified way to separate real Claude competency from self-reported case studies.

40 production agents shipped. Our writing, pricing, and scoping is calibrated on 40 deployments between 2024 and 2026, not on marketing theory. Median cost-per-request across the fleet: $0.008–$0.041 for ops agents. First-call tool-use success rate: 96.1% on Claude Sonnet 4.6, 94.3% on GPT-5.4. P95 latency range: 6–22 seconds depending on workflow. Those are the numbers we'd want our own proposals to be checked against, and they're the numbers we quote in scoping calls instead of adjectives.

The $1,500 one-week audit, refundable against the build. We start every engagement above $20k with a one-week audit ($1,500, fixed-price). At the end of the week you have a scoping document, an architecture recommendation, and a fixed-price build proposal. If you sign the build, the $1,500 is credited against the build. If the audit concludes "this workflow doesn't need an agent, it needs a Zapier scenario," we say so, you pay $1,500, and we don't get the build. Roughly 15% of our audit conversations end that way. That's the number that matters; the rest is marketing.

Two pillars we publish on. We publish production postmortems on the Claude Agent SDK production playbook (five patterns we ship) and the OpenAI Agent Builder vs Claude Agent SDK comparison (the decision framework for picking either platform). Both are written for engineers, not for buyers — they're where you'd look to check whether we can actually do the work before you talk to us.

An eval harness shipped with every build. Every engagement above $20k ships with a golden-trace eval suite (20–50 cases, tailored to the workflow), a drift-detection job, and a dashboard tuned to your SLAs. That's the thing that turns model updates from a fire drill into a Tuesday task, and it's what separates a delivered agent from a delivered project. If an agency ships a production agent without an eval harness, the agent is a regression waiting to fire.

A clean exit clause. Our SOWs have a termination-for-cause clause that hands you the full code, the eval suite, the dashboards, and a 2-week transition — no hostage-taking on deliverables. If the engagement isn't working, you should be able to walk with the work you paid for. A studio that won't agree to a clean exit clause is telling you something about how they expect the engagement to go.

None of these differentiators are individually rare — there are other archetype-1 studios that ship all five. Collectively, they're the combination we'd look for if we were the buyer on the other side of the table.

What to ask an agency before you sign

Ten questions. Run them on every agency on your shortlist. The answers don't need to be identical across shops — they do need to be unambiguous, and any ambiguity on these ten is a signal.

1. Can I see two production case studies with dollar impact and a measurement window?

You're looking for "X customer, Y workflow, $Z saved per year, measured over N months." If the answer is three adjectives and no number, the case study is either marketing or early-stage. An archetype-1 studio will have at least two on request, sometimes under NDA but shareable under mutual NDA.

2. Which agent SDK or framework are you planning to use, and why?

You want to hear "Claude Agent SDK because X, OpenAI Agents SDK because Y, LangGraph because Z" — an opinion rooted in the shape of your workflow. If the answer is "we'll use whatever model works best" or only "Make / n8n," you're in archetype 3 or 4. Good studios have a default opinion and can articulate when they'd override it.

3. What does your eval harness look like?

A real answer names the structure — number of golden cases, how they're generated from real traffic, how regressions are caught, what the dashboard looks like. If the answer is vague or the agency asks you to explain what you mean by "eval harness," they don't have one.

4. Who specifically will be writing the code?

Ask for the engineer by name. Ask for a 20-minute technical call with them before signing. If the agency can't or won't set that call, you're in archetype 6 (reseller) or a stretched archetype 5 (founder-only, won't scale). An archetype-1 studio sets the call gladly, because the engineer selling the work is usually the engineer shipping it.

5. What happens if the model provider releases a new model during the project?

Right answer: "The eval harness catches regressions on golden traces, so swapping models is a scoped task, usually 1–3 days. Here's our process." Wrong answer: "We'll re-prompt it, it'll be fine." Model updates are frequent — Claude Opus 4.7 shipped April 16, 2026, and OpenAI's Agents SDK shipped a major sandbox-execution + model-native-harness update one day earlier on April 15 with native support for Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop, and Vercel⁴ — and shops that don't have an eval process will quietly absorb regressions into your production traffic.

6. What does handoff look like at the end of the engagement?

Right answer: "Code in your repo, 30-day hypercare window, runbook for your on-call, eval suite you own, two-week transition meeting cadence." Wrong answer: "We'll stay on retainer, we're your AI team." The second answer means the agency has designed the engagement so you can't leave.

7. What does your SOW's termination-for-cause clause look like?

Right answer: a clean clause that hands you the code, eval suite, dashboards, and a defined transition. Wrong answer: "We haven't had to terminate an engagement yet." Every SOW should have the clause before anyone has a reason to use it.

8. What SLAs do you ship with?

Right answer: p95 latency, per-request cost ceiling, first-call tool success floor, error-rate ceiling — each with a numeric commitment. Wrong answer: "We target high reliability." An agency that can't commit to numeric SLAs either hasn't measured them or is planning to hand you whatever the model happens to return on launch day.

9. How do you price, and why?

Right answer: a specific pricing model (fixed-price, retainer, audit-to-build) with a clear rationale for why it fits the scope. Wrong answer: "We'll customize a quote for your needs" with no further specifics. Pricing clarity is a proxy for operational clarity.

10. Would you recommend against a custom build in any case, and under what conditions?

Right answer: "Yes — if the workflow fits a SaaS product under $2k/month, we'll recommend the SaaS and decline the engagement. Here's the last time we did that." Wrong answer: "We can customize any workflow." The second answer is the confirmation that you're talking to archetype 4, 5, or 6. An archetype-1 studio turns down business routinely, because not every workflow needs a custom build.

A note on discovery calls. The discovery call is where archetypes reveal themselves. An archetype-1 studio will spend most of the call asking about your data, systems, edge cases, and metrics. An archetype-4 or -5 studio will spend most of the call asking about your revenue, your headcount, and your "pain points" in sales terms. The questions are tell. You often don't even need to ask the ten above to sort the shops; the shop's own questions will do 80% of the filtering for you.

FAQ

The frontmatter FAQ covers the questions we field most often on intake — pricing bands, timelines, in-house vs. agency, proposal anatomy, eval harnesses, Claude Partner Network relevance, and ROI measurement. If your question isn't there and you want a grounded answer, book a strategy call.

The short version of every FAQ. The six-archetype framing is the decision. Once you've placed the agency into an archetype, the pricing model, the stack, the proposal shape, and the ROI math mostly follow. Pick archetype 1 for ops production work. Pick archetype 2 only if you've already committed to the platform. Treat archetypes 3–6 as specific fits or passes, depending on your workflow shape and risk tolerance.

Further reading

If you're evaluating a specific engagement and want the engineering depth behind our recommendations:

• Claude Agent SDK in production: a studio's playbook — the five patterns we use when Anthropic's SDK is the right call, including the subagent decomposition and eval-gated deployment patterns that define archetype-1 work.
• OpenAI Agent Builder vs Claude Agent SDK: a studio's decision framework — head-to-head on the two platforms we ship on, including cost and latency benchmarks from 40 agents.

If you want us to build or audit one of these with you, book a strategy call — our one-week audit is $1,500 and refundable against a build if you move forward.