What an AI Automation Consultant Actually Does, and What It Costs

Your company pays for AI already. There are ChatGPT seats on the card, someone expensed Claude, the sales team has an AI notetaker, and a developer swears Copilot has changed their life. Add it up and it is a real line item.

Now look at Tuesday. The invoices still get keyed in by hand. Someone still triages the shared inbox for forty minutes every morning. The weekly report still eats a person's Friday. The same three exceptions still get escalated to the same manager. Nothing about how the work actually moves has changed.

That gap is the whole business of AI automation consulting in 2026, and it is worth being precise about what closes it. This is a straight answer to what the job involves, what it costs, what drives that number, and how to tell someone who will ship from someone who will deliver a slide deck and an invoice.

The gap is not access to AI. It is integration.

Every analyst writing about small and mid-sized businesses this year lands on the same finding, and it matches what we see in the field: the constraint is no longer tool access. The models are good, cheap, and available to everyone including your competitors. The constraint is that the tools are built for individual productivity and your problems are operational.

A ChatGPT seat helps the person holding it, one prompt at a time, as long as they remember to open it. It does not watch your inbox at 6am. It does not reconcile last month's statements. It does not move data between your CRM, your accounting system, and the spreadsheet that finance actually trusts. Those things require software that runs whether or not anyone remembers it exists, and building that is engineering work.

This is also why the "we tried AI and it didn't stick" story is so common. A subscription is not a strategy. Without a workflow, guardrails, and someone who owns it, the tools get opened twice and forgotten, and the organisation concludes AI does not work for them. What did not work was the deployment.

What the job actually is

Strip away the positioning and an AI automation engagement is three pieces of work, in order.

Find. Somebody sits with your operations and produces a process map: how the work moves, where the handoffs are, which steps are pure judgment and which are rules wearing a costume. Out of that comes a ranked list of automation candidates with an estimate of what each returns, in hours and in errors that stop happening. This is the part buyers most often skip and most often regret skipping. Deploy tools before you have decision rules and you get fragmented adoption: six pilots, no compounding, and nobody able to say what any of it earned.

Build. One workflow at a time, highest payback first. Built against your real cases, including the ugly ones, then wired into your actual systems with scoped credentials and a human path for everything the AI should not decide alone. Shipped into daily use with the people who will run it.

Embed. Training in plain English on the workflows you now own, monitoring so you find out when something breaks before your customer does, and documentation the next engineer can read. Then the next workflow.

Notice what is not on that list: a maturity model, a two-day workshop, or a report recommending you consider a broader transformation programme. The measure of this work is software running in production and hours that come back to your team.

Agents versus chatbots, in one paragraph

The word "agent" is doing enormous marketing work right now, so here is the distinction that matters commercially. A chatbot answers questions. An agent does work. A chatbot tells your AP clerk what the payment terms usually are. An agent reads the invoice, matches it to the purchase order, checks the terms, updates the ledger, and escalates the one that does not reconcile to a human with the discrepancy already highlighted.

That difference is why agents return hours instead of answers, and also why they cannot be installed like an app. An agent has permissions. It touches money paths and customer data. It needs to know what it must never do alone. That is engineering and it is the reason this work is priced as a project rather than a subscription.

Where AI actually pays first

Agents earn their keep on repetitive judgment calls with clear rules and digital inputs. Across the market, and in our own engagements, the same handful of workloads come up first.

• Document processing. Invoices, receipts, contracts, forms: read, validated, routed. The most reliable payback in the set, because the input is structured enough to check and the manual version is pure cost.
• Customer communication. Inbox triage, drafted replies, follow-ups that actually go out, support workflows with human sign-off where the stakes require it.
• Operational reporting. The weekly numbers assembling themselves from your systems and landing in your inbox before anyone asks. Dashboards that stay current without a person feeding them.
• Sales and proposals. Lead research, CRM hygiene, meeting notes turned into follow-ups, proposals that start 80% drafted instead of from a blank page.
• Finance workflows. Reconciliation, expense processing, chasing AP and AR, the month-end grind, with an auditable trail and a human on the exceptions.
• Internal knowledge. Your policies, project history, and documents made answerable, so staff stop re-asking and re-searching for what the company already knows.

If a vendor's pitch does not start by naming which of these applies to you, they are selling you the category rather than the outcome.

What it costs

Published pricing in this market is genuinely wide, because "an AI automation project" spans a script that files your receipts and a programme that reorganises a department. Here are the bands that hold up across 2026 sources, so you can place any quote you receive.

• Assessment or roadmap: roughly $2,000 to $25,000 depending on depth, typically 2 to 8 weeks. Produces the process map and the ranked plan, no working software.
• Pilot on one workflow: $5,000 to $25,000, usually 4 to 8 weeks. Exists to prove the return before anyone commits further.
• Production deployment with real integration: $25,000 to $100,000. This is where most serious mid-market work lands.
• Enterprise programmes: $100,000 to $500,000 and up, over 6 to 18 months.
• Hourly, if someone bills that way: offshore agencies from about $25, boutiques with senior engineers delivering directly at $150 to $300, brand-name firms well beyond that.

Our own floor is lower than most of those bands, deliberately. A bounded integration, meaning one workflow wired into one or two systems, starts at $3,500, because the fastest way to find out whether AI pays in your business is to automate one real thing and count the hours. From there, price scales with the size of the company and the depth of the integration, and there is no fixed ceiling: the scope is whatever your operations justify.

The uncomfortable truth about this market is that the same engagement can be quoted at $50,000 or $500,000 depending on who is quoting. The difference is usually not quality of outcome. It is overhead, layers of account management, and whether the person who understood your problem is the same person who writes the code.

What actually drives the number

When you get a quote, these are the five variables behind it. Any consultant should be able to walk you through them without hedging.

• How many workflows. The obvious one, and the easiest to control. Start with one.
• How many systems it touches. One system is a script. Four systems, two of which have no usable API and one of which is a spreadsheet on a shared drive, is a project. Integration surface drives cost more than model choice ever will.
• How much judgment the AI exercises. Extracting a total from an invoice is cheap to get right. Deciding which invoices to pay is not, because being wrong costs money, so it needs review paths, thresholds, and audit trails.
• What the failure mode costs. An automation that occasionally mislabels an internal ticket needs less engineering than one that emails your customers. Stakes buy safety work, and safety work is real hours.
• Data condition. If your records are clean and accessible, the build starts on day one. If they live in three places with different customer IDs, someone is reconciling that first, and pretending otherwise is how projects slip.

The part nobody sells you: security

Every automation worth building touches something sensitive. Credentials, customer records, money paths, or all three. An agent with API keys is a user who never sleeps, never asks permission, and does exactly what its instructions say, including instructions that arrive inside the content it is reading.

Ask any prospective consultant these four questions and listen for specifics rather than reassurance.

• What credentials will this run with, and what is the smallest set of permissions that works?
• Where does our data go, which vendors see it, and what do their terms say about training on it?
• Which decisions require a human, and what does the agent do when it is not confident?
• What happens when it fails at 2am, and who finds out?

If the answers are vague, you are being sold a demo. This is the reason we think AI adoption belongs with people who also do security review work: the failure modes of a badly wired agent are not theoretical, and the cheapest time to design them out is before the thing is built.

How to tell a builder from a slide deck

Five signals, in rough order of usefulness.

• They ask about your operations before they talk about models. Which model you use is close to the least interesting decision in the project.
• They have run something in production for years, not months. Anyone can demo. Keeping AI working against real data, real staff, and real edge cases is a different discipline, and it is the one you are buying.
• They will tell you when not to automate something. A consultant who finds an AI use case in every corner of your business is describing their revenue model, not your operations.
• The quote has reasoning attached. You should be able to see which of the five cost drivers above produced the number.
• They hand over documentation and training. If the automation is only maintainable by the people who built it, you have not bought an asset, you have bought a dependency.

What good looks like, concretely

We have run this pattern since 2017 on Receipt Rewards, the OCR document-processing platform behind national promotions for Unilever, Knorr, Hellmann's, Scotch, Fandango, and Novamex. Roughly 100,000 receipts read, validated, and adjudicated by AI, across 3 to 6 national campaigns a year, for a client who is not a technology company and does not want to think about any of this.

The interesting part is not the model. It is everything around it: what happens when the photo is blurry, when the receipt is a fraud attempt, when a promotion changes mid-flight, when volume spikes on launch day. Automation that survives contact with the real world is mostly the handling of cases the demo never shows you. That is what you are actually paying an engineer to have seen before.

When you should not hire anyone

If the repetitive work in your business adds up to a few hours a week, buy the tools, spend an afternoon with your team on how to use them, and get on with your life. Automation has fixed costs, and below a certain volume the honest advice is that it will not pay.

The maths changes when a task is measured in days per month, when errors in it cost money or customers, when the work scales with revenue so growth means hiring, or when the person doing it is expensive and would rather be doing something else. That is when an automation that runs unattended stops being a nice idea and starts being the cheapest employee you will ever hire.

Bring the workflow your team hates most. In a free 30-minute call we will tell you whether AI can take it off their hands, roughly what that would cost, and what it would return. Straight answer either way, including "not worth it".