Insights / AI
AI Automation Consultant: What the Role Should (and Should Not) Do
An automation consultant should leave you with a map, a two-week pilot, and a date to stop if it fails. An unsupervised agent on day one is a demo, not a system.
What an AI automation consultant is hired to design
An AI automation consultant is hired to design a loop, not to perform a model. The loop has a trigger, a draft or an action, a rule for when a human must look, and a place the result is stored. If they cannot draw that on one page, they are not consulting. They are demonstrating.
In a 10–40 person company, automation is not a department and not a platform war. It is the decision to stop typing the same reply, chasing the same invoice, or rebuilding the same weekly report. The consultant’s job is to pick that loop, keep it small, and refuse the rest for now.
I say this as someone who builds software at BlocksGenie and runs operations at VPSGrid: the companies that get value are boring on purpose. They automate the queue. They do not automate the exception. They do not start with an agent (software that takes actions — emails, updates records — without a person in the loop each time) that has permission to speak for the brand.
The deliverable is a written workflow, not a demo
Start with the work as it exists, not as a vendor wishes it existed. Who starts it. What they look at. Where they copy from. What “done” means. What they do when the input is wrong. If that map is missing, every tool will look equally promising and equally useless.
Write the map in the language of the team, not the language of the model. “Ticket arrives, a support person reads it, searches three docs, drafts, a senior reviews if refund or legal, then sends” is a map. “We will deploy a multi-agent graph” is a press release.
Then mark the steps that are volume, the steps that are judgment, and the steps that are irreversible. Volume can often be drafted by a model. Judgment stays human. Irreversible — refunds, legal, medical, public statements — stays human until you have a long, boring success record. That marking is the design.
A good consultant will also name the system of record — the software where the official copy of the work lives. Helpdesk, CRM (contacts and deal stages), CMS (content management system, such as WordPress), billing. The automation must write the result back there. If the result lives only in a chat window, you have not automated. You have added a tab. ChatGPT can be a step in the loop, and depending on plan and settings, apps or connectors may let it reach some of those systems. It still does not replace the loop: someone has to own the permissions, the review rule, and the system of record.
Existing tool versus custom work — decide it on purpose
Three honest options, in order of cost and regret.
- Built-in feature. The helpdesk already drafts replies. The CRM already summarises calls. The CMS already suggests titles. Turn it on, write the review rule, measure. For many small firms, this is a sensible place to stay for months before anything custom.
- Chat-class tool plus a human integrator. A person pastes, edits, and files. Ugly. Cheap. Reversible. Good enough for SOP drafts (written how-tos), research briefs, and reply templates you will later store in the built-in tool.
- Custom. An integration, a script, an agent with permissions, a write-back to billing or the CMS. Use this when the step must run without a tab, or the data cannot be pasted. Custom is a project. It needs a spec and an owner.
The consultant should recommend the cheapest option that meets the metric. If they skip to custom, ask what built-in option they rejected and why. If they cannot name one, they skipped the work.
WordPress, used as a CMS, is a common example. A plugin that drafts meta descriptions is close to a built-in feature. A custom agent that publishes posts unsupervised is a different risk class. Do not let a demo collapse those two into “we’ll AI your CMS.”
How to run a two-week pilot with a pass/fail test
A pilot is not a prototype you are emotionally attached to. It is a test with a date. Two weeks is long enough to see volume and short enough that a failure is cheap.
Define pass/fail before day one. Illustrative examples — targets you set, not benchmarks: first-response draft time drops by 40 percent with no increase in factual corrections; invoice chasers go out within 24 hours of due date with a human still approving the first three exceptions. If you cannot write the sentence, you are not ready to start.
Staff the review. Someone checks a sample every day for two weeks. Not “when we have time.” A consultant who cannot name that person should refuse the work. Automation without a reviewer is how you scale a wrong reply.
End with keep, change, or kill — written. Change means the prompt or the rule, not a new platform. Kill means you stop. Keep means you train a second person and only then discuss custom. The five-question audit is how you chose the candidate. The pilot is how you find out if you were right.
| Pilot element | Required | What goes wrong if it is missing |
|---|---|---|
| Named process | One sentence | Scope creeps into “all of support” |
| Metric | One number, day 14 | You debate feelings |
| Reviewer | A person with calendar time | Errors reach customers |
| Kill date | On the calendar | A process keeps running with no owner and no metric |
| System of record | Where the output is saved | Work trapped in chat |
What an AI automation consultant should not do
Should not dump unsupervised agents as step one. An agent that emails customers, posts in public, or moves money is a production system. Production systems need a boring pilot behind a human send button first. Starting with autonomy is a demo for the buyer’s ego.
Should not sell a new data warehouse as the first automation. Warehouses are months. First automations are weeks. If the consultant needs a warehouse to start, they are solving their architecture preference, not your queue.
Should not install a twelve-tool stack. Each tool is a login, a bill, a failure mode, and a person who must care. One loop, one or two tools, one reviewer. Stack comes later, if the metric demands it.
Should not become your operator. If the workflow only runs when they are in Slack, you have a contractor with a clever title. Transfer the loop in week two: your accounts, your prompts, your reviewer. An AI consultant for small business who will not leave the keys is not done.
How to brief them in one page
One page. Process name. Volume per week. Tools you already pay for. Five redacted examples. The last time it failed and what it cost. Who will own it next month. The metric for day 14. Whether you are about to spend on a developer.
Do not send a seventy-page SOP dump without a cap. Pre-session review should be limited. If they accept unlimited reading at a few hundred dollars, they will resent the work or skim it. Neither helps you.
Bring the person who does the work, not only the founder. Founders describe the process as they wish it were. Operators describe the exceptions. Automation dies on exceptions. The consultant needs to hear them in the room.
State your constraint on autonomy in the brief: “Nothing sends to a customer without a human click for the first thirty days.” Good consultants will agree. The ones who argue are selling a demo.
What “automation” means in a 10–40 person company
It means a repetitive path with a named owner, not a replacement for the people who hold customer relationships. A 15-person professional-services firm can automate first-draft tickets, meeting summaries, and invoice reminders. It should not automate the recovery conversation after a missed deadline.
It means write-back to systems you already run — saving the result into the software you already use. If billing lives in one place and the CMS in another, the loop must respect that. Do not add a third system of record because the vendor’s agent “needs a knowledge graph.”
It means reversible first. Reply templates, drafts, reminders, internal summaries. Irreversible later, and only with a log. Refunds, public posts, legal language, access changes — those are last, not first.
It means a kill date. Small companies accumulate unused automations and unused prompts the same way they accumulate unused subscriptions. The consultant should put the kill date on the calendar in the room. If they will not, you put it on.
How this sits next to hiring and tools
If you do not know which process, you do not need an automation specialist yet. You need diagnosis. Implementation, development, licences, and ongoing Slack are not included in these sessions.
| Package | Price | What you get | Best for | Not included |
|---|---|---|---|---|
| AI Discovery Call | $149 · 45 min | Direction in the session | You only have questions | Written summary, build, licences, Slack after |
| AI Business Audit (featured) | $299 · 60 min + post-session summary | Which process to automate first, in writing | You need one process named | Development, licences, ongoing Slack |
| AI Implementation Blueprint | $599 · written roadmap | A sequence a developer can follow | You know the process and need it in writing | Coding, deployment, licences, ongoing management |
If the work is still “paste into a chat window,” you may only need a rollout of ChatGPT for small business, not a consultant at all. Hire when you are about to connect systems, buy a platform, or give a model permission to act.
You can use this article this week without a booking: draw one loop on a page, mark volume versus judgment versus irreversible, and pick a two-week metric. If you cannot, do not hire anyone to automate it. If you can, you already know what to brief.
A filled brief for a 15-person company
Process: first-response drafts for support tickets that are not refunds and not legal. Volume: about thirty-five a week. Tools: helpdesk plus a ChatGPT-class workspace. Examples: five redacted tickets, one gold-standard reply (your last good reply). Last failure: a wrong SLA promise (service-level agreement — the response time you promised) went out because a template was stale; it cost a credit and an hour of a senior. Owner: support lead, thirty minutes at 4 p.m. daily. Metric: time to first draft, and factual corrections per twenty tickets, reviewed on day 14. Autonomy constraint: no send without a human click for thirty days. Developer spend: none planned until the metric exists.
That brief is enough for a $149 Discovery Call or a $299 Audit. It is not enough for a custom agent. Anyone who reads it and quotes a platform licence has not read it. Send it anyway. The wrong vendors show they are the wrong fit.
If you cannot fill those fields, you are not briefing an automation consultant. You are asking someone to invent your operation in an hour. They will invent something that looks like their last client. You will then spend a quarter undoing it.
What usually goes wrong in week two
The reviewer disappears. Output still generates. Nobody samples it. By day ten the loop is fiction. Put the review on a recurring calendar invite before day one, or do not start.
Exceptions flood the happy path. Staff start stuffing edge cases into the same prompt because it is open. The metric collapses. The fix is routing, not a smarter model: a short list of types the loop may touch, and a button that means “this is not a template ticket.”
The consultant is still in the channel, quietly fixing prompts. That feels like service. It is a transfer failure. Week two is when your owner makes the edits. If they cannot, the loop is too fragile to keep. Kill it or simplify it. Do not retain the consultant as a hidden operator on Slack.
A new tool appears “just for this pilot.” That is going back to shopping. Built-in feature or the workspace you already pay for. If those cannot run the two weeks, you chose the wrong process. Go back to what to automate first and score again. Automation consulting is not a licence to add more software.
Questions people ask
What does an AI automation consultant do?
They map a repetitive process as it exists, choose existing tools versus custom work, and sequence a two-week pilot with a pass/fail metric and a kill date. They should not dump an autonomous agent on you as step one.
What should an AI automation consultant not do?
Unsupervised customer email on day one, a new data warehouse as the first project, a twelve-tool stack, or uncapped Slack. Those are demos, long research projects, or retainers without edges.
Do I need an automation consultant or a developer?
Consultant first if you do not know which process or whether a built-in feature is enough. Developer after the process is named and the tool-versus-build choice is made. Paying build rates for discovery is the expensive order.
What does automation mean in a 10–40 person company?
A named loop: trigger, draft or action, human review where it matters, write-back to the system you already use. It does not mean replacing judgment, the closer, or the person who calms an angry customer.
How do I brief an automation consultant?
Process name, volume, tools, five redacted examples, the last failure and its cost, who will own the output next month, and a metric you will look at on day 14. A slogan is not a brief.
How much should I pay before a build?
Direction at $149, diagnosis at $299 with a summary, or a $599 written blueprint if implementation is imminent. Development is a separate contract. Hourly with no cap is the trap.
When is a two-week pilot a failure?
When you cannot show a number, when nobody reviewed the output, or when the workflow only works if the consultant is in the channel. Kill it. Killing a bad pilot is a successful engagement.