What service businesses are missing from the AI agent conversation

Open any AI agent demo on the timeline and it is one of two things.

It is a consumer chatbot. Someone asks an agent to plan a trip to Tokyo and the agent books eight flights and the demo ends. Or it is an enterprise pilot. A Fortune 500 PR release with a logo grid that says "AI agents in production at the bank." The pilot covers one workflow, took eighteen months, and required a six-figure professional services engagement to launch.

The segment between those two extremes is empty.

It is the segment where AI agents pay back inside a month. Where the buyer can decide on a call. Where the work the agent replaces is judgment-shaped and repeated and currently being done by a human with a spreadsheet. Where the runtime cost is small enough that the agent is profitable on day one.

That segment is small service businesses. Agencies, consultancies, accountants, lawyers, designers, recruiters, anyone selling expertise as a service to other businesses. It is also where I sit, and it is the segment the agent market is shipping nothing for.

This post is what the missing agent looks like, what nobody is building, and what I see clearly from running a small services firm that uses agents every day.

The shape of work in a services business

Sit next to a partner at a small consultancy for a week and write down what they do.

Most of their day is judgment work that requires reading something, deciding something, and writing several things in several places. A new lead came in. They read the LinkedIn. They check whether the company is a fit. They draft a reply. They update the CRM. They Slack the team. A discovery call just happened. They read the transcript. They write a recap. They update three fields on the deal. They draft a proposal section. They flag a follow-up. A monthly customer review is coming up. They open the last four months of activity, write a summary, draft a renewal pitch, send it.

None of this is hard work. Most of it is not even unique work, because the partner is doing the same shape of thing they did for the customer before and the customer before that. It is repeated judgment, shaped by the partner's experience, that lives in their head and their inbox and their Notion.

This is exactly the shape an agent absorbs well. The model can read the LinkedIn. It can hold the deal context. It can write the recap. It can update the CRM through MCP. It can draft the renewal pitch. The judgment part used to require the human, and as of 2025 it does not. The agent does not replace the partner. It does the partner's first draft so the partner can edit instead of write from scratch.

You would think there would be an agent on every partner's desk by now.

There is not. And the reason is not the agent. It is who the market has been shipping agents for.

The two markets the agent industry is actually building for

The first market is consumers. ChatGPT, Claude.ai, Gemini, every other chatbot you can name. The agent is a chat box. The user types. The agent answers. The value is "what would have taken me twenty minutes I now do in two." That is real value, and it is also the value of a smart calculator, not of a system that runs the work of a business.

The second market is large enterprises. Salesforce Einstein. HubSpot Breeze. ServiceNow Now Assist. Microsoft Copilot. The pattern is the same: a year-long pilot, a six-figure committee, a workflow scoped to one team in one region, and a Q3 press release. The agent runs on the customer's data, fronted by the vendor's UI, billed per seat at enterprise rates. The buyer is the CIO. The check is six figures because the procurement cycle requires it.

Neither of these is what a thirty-person agency needs.

A thirty-person agency does not have a CIO. They have a founder, an ops person if they are lucky, and a partner who said "we should use AI for this." They cannot run an eighteen-month pilot because they do not have eighteen months. They cannot pay six figures for one workflow because their margin per engagement is not six figures. They also cannot use the consumer chatbot for production work, because the chatbot does not write back to their CRM, does not run on a schedule, does not have access to the customer's call transcripts, and does not know their team's voice.

So the partner opens ChatGPT, pastes the LinkedIn into the prompt, gets a draft, edits it, pastes it back into the CRM by hand, and goes back to wishing there was a better way. The agent industry has shipped nothing for them.

What the missing agent actually looks like

The agent the small services business needs has a specific shape, and you can almost write it as a checklist.

It runs on a schedule, not as a chat. Monday at 8 a.m. the agent triages the deal pipeline and writes a one-line note on every stuck deal. Friday at 4 p.m. the agent reads the week's call transcripts and writes the meeting recap to each deal. The partner does not summon the agent. The agent shows up.

It writes back to the system of record. The output is not a chat reply. It is a CRM update, a Notion page, a Slack message, an email draft. The agent's job is to leave the workspace in a better state than it found it, not to produce a response in a chat window.

It is shaped to the team's specific workflow, not to a generic feature. "Summarize this call" is a feature. "Read the call transcript, find the three things the buyer cares about most, write a follow-up draft in our team's voice, and tag the deal with a competitor mention if one came up" is a workflow. The first ships in every CRM by default. The second has to be written by someone who understands the team.

It costs less per month than the work it replaces. A partner-hour costs $200 minimum. An agent that runs once a day and saves a partner an hour a day is worth $4,000 a month and costs $30 in model usage. The economics are not subtle. The agent pays for itself on the first run.

It is owned by the customer, not by the vendor. The agent's instructions live in a folder the customer can read. The runtime is the customer's account, not the vendor's server. If the vendor disappears, the agent keeps running. This is the part most builders get wrong, and it is the part the careful buyer cares about most.

None of those five characteristics describe a consumer chatbot. None of them describe an enterprise pilot either. They describe a third kind of product that the industry has not really built yet.

Why nobody is shipping it

Three reasons, and they compound.

The unit economics look unsexy from a venture perspective. A small services business pays $300 to $3,000 a month for an agent. There are maybe a hundred thousand of them in the markets where the agent makes sense. Top of the market by revenue is a few billion dollars. That is a good business and a terrible venture bet. Vendors with venture money optimize for either the consumer billion-user category or the enterprise six-figure-per-customer category, both of which scale to the trillion-dollar outcome the fund needs.

The build is custom-shaped per customer, which scares software companies. Each services business has its own pipeline stages, its own qualification criteria, its own deal review format, its own renewal cadence. The agent has to be configured for those specifics. Software companies that grew up on "ship one product, sell to a million customers" are not built to ship one product per customer. They route this work to the systems integrator partner channel, and the partners price it at enterprise rates, and the small services business is back to square one.

The buyers do not know what to ask for. A founder of a thirty-person agency does not write a procurement document for an AI agent. They do not know what a workflow looks like in instruction form. They cannot tell a vendor what good looks like, which means even when a vendor wants to ship to them, the spec is unclear and the engagement falls apart.

The result is that the market is structurally underserved. Not because the technology is missing. Not because the demand is missing. Because the shape of the buyer does not match the shape of any vendor who currently has agents to sell.

What does work, and who is doing it

The agent that fits the small services business looks less like a product and more like a practice.

The practice is small. One to five people. They have a specific domain (Attio implementations, in my case; could be Klaviyo for ecommerce, could be specific legal workflows, could be tax for SaaS founders). They write agents as configuration on top of an LLM provider's runtime, not on their own server. They sell the configuration once and operate the workspace forever. The runtime cost lives on the customer's account. The practice's margin is in the build and the operation, not in the rent.

This pattern has shown up in the last twelve months across small categories. Implementation partners for specific CRMs. Vertical-specific consultancies that ship agents alongside the audit. Operators-turned-builders who shipped agents on top of one workflow they understood deeply.

None of these are venture-backed. None of them are running pilot programs with logo grids. Most of them are profitable in month one because the build is one engagement and the operation is on the customer's runtime.

I run one of these. We ship Attio implementations and ship agents on top. Six agents are running in customer workspaces today. None of them sit on a server I own. The economics work because I am not infrastructure. The customer's Claude Code subscription is the runtime. My contribution is the markdown that encodes the work. The math is what makes this a real business and not a hobby.

What is missing from the AI agent conversation is the recognition that this shape of business is the actual production deployment of agents in 2026. Not the consumer demo. Not the enterprise pilot. The hundred thousand small services firms, each running a handful of custom agents shaped by a small practice that knows their domain.

What I think happens next

Three things, and they are already underway.

The consumer chatbots will keep growing because they are useful and the distribution is enormous. They will not become production agents for service businesses. They will stay smart calculators that one person uses one prompt at a time.

The enterprise vendors will keep shipping their pilot programs because their go-to-market requires it. Some of those agents will eventually be good. The customer who pays for them will pay enterprise rates.

The small practice layer will grow underneath, mostly invisibly, because the people doing the work are not loud and the customers buying it do not write LinkedIn posts about their CRM. This is the layer where AI agents are going to do most of their work in 2026 and 2027. Most of the value will accrue to operators who understand a domain and ship agents shaped to it, not to vendors who ship a generic AI feature with a multi-tenant runtime.

If you are a partner at a small services firm and you have been waiting for the agent that solves your specific workflow, the agent is not coming from the vendor whose logo is on your laptop. It is coming from a one-person practice in your space who already understands what you do.

The market will eventually catch up. Until then, the work that gets done is going to come from operators who got tired of waiting and built it themselves.

If you run a services business and the work you wish you could hand to an agent is sitting in a Notion doc somewhere, I am happy to talk about what that agent could look like. I write more about the practice I run at craftt.io, and there is a longer post on why an AI agent matters more than a CRM feature that pairs with this one.