How much does AI consulting cost, and what should it include?

A good AI consultant finds a few specific tasks in your business where AI can save time or prevent mistakes, builds something that does that work on your own data, and checks with you whether it helped. The result is a working tool your team uses. A strategy document can be part of it, but if the work ends there, nothing in your business has changed yet.

Whatever the size of the project, you should come away with four things.

• A named task, specific enough that anyone on your team would recognize it, like sorting incoming quote requests or answering common questions by email
• A working tool that runs on your real information, tested by the people who'll use it
• A way to measure whether it helped, agreed on before the work starts, such as the time the task takes or how often something gets missed
• A person responsible for checking what the tool produces, at a clear point before it reaches a customer or your records

This kind of AI work is about how things get done inside your business. If you're asking how your business shows up in ChatGPT or Google's AI Overviews, that's a website question, and getting found in AI search covers it.

We run AI work in four steps. The second one ends with your decision about whether to go ahead, so nothing gets built until you've seen what it would take.

Assessment

We start by looking at how work actually moves through your business. Where does information come in, who handles it, what gets typed twice, and what gets missed? The goal is a short list of tasks that take the most time or cause the most mistakes. This part doesn't involve any AI yet.

Recommendation

We recommend where to start and what it would take, so you can decide whether to go ahead. That's usually a small number of narrow tasks, each with the reason it's worth doing. Sometimes a simple form, a rule in software you already have, or a change in who does what would handle a task better. When AI isn't the right answer, we'll tell you.

Implementation

We set up what we agreed on, run it against your real data, and adjust it based on your feedback. Testing on real examples matters, because a tool that works on a few clean samples can still fail on the messy ones your team sees every day. This is also where the review step gets built in, so a person can check what the tool produces before it reaches a customer.

Review

We go over the results with you, using the measure agreed at the start, and decide together what to change next. That can mean adjusting the tool or moving on to the next task on the list.

AI works best on narrow, repeated tasks with a clear input and a clear output, where a person can check the result faster than doing the work from scratch. That last part is a useful test. If checking the result takes as long as doing the task, the tool isn't saving anyone time.

Where it tends to help

• Reading incoming emails, forms, or attachments and turning them into clean records in the system you already use
• Drafting replies to common questions, which a person reads and approves before they go out
• Searching across your own manuals, past quotes, and policies, and showing where each answer came from
• Flagging the few exceptions in a long list, like invoices that don't match an order, so someone only looks at the ones that matter
• Summarizing long notes or call recordings into the fields your team fills in anyway
• Moving information between tools that aren't connected, as a workflow automation
• Rough concept images for talking through an idea before paying for final design or photography

Where it usually doesn't

• Decisions that need judgment and someone accountable for them, like pricing a job, hiring, or telling a customer no. AI can lay out the options, and a person should decide.
• Anything where one wrong answer is expensive and nobody would catch it, like the final numbers on a quote or wording with legal weight.
• Tasks that come up a few times a year. Setting up and checking a tool takes longer than doing them by hand.
• Questions the tool has no source for. AI models can produce answers that are confident and factually wrong, so a tool that isn't given your real information will guess.
• Problems that are really about messy data or a missing process. If the numbers in a spreadsheet are wrong, an AI tool will repeat them.

That last one is worth checking first. Sometimes the right fix is ordinary software, like a form that feeds one database or a dashboard everyone works from, and our post on when to build custom software explains how to tell.

The cost follows the work, so the same things that make a task harder to automate make it cost more to build. Running the tool afterward is a separate cost, and a proposal should cover both.

Building it

• How many tasks are in scope. One narrow task is a smaller job than several at once, which is one reason to start with one.
• Where your information lives. If it's already in one clean system, the work is simpler. If it's spread across email, spreadsheets, and someone's memory, gathering and cleaning it takes longer.
• What it has to connect to. A tool that works inside your existing software takes more setup than one that stands alone, and some end up as custom software in their own right.
• How much a mistake would cost. Anything touching money, records, or customer communication needs review steps built in, and that adds work.
• How many unusual cases there are. A task with lots of exceptions takes more rounds of testing on real examples before it's reliable.

Running it after launch

Most AI tools cost something every month once they're live. A proposal should name each of these that applies.

• Usage fees. Tools built on an AI provider's service are usually billed by the amount of text the model reads and writes, so the bill follows how much you use it. More capable models cost more per use, and simple tasks can often run on a smaller one.
• Subscriptions. Chat tools your staff use directly are usually priced per person, per month.
• Other services in the chain, like the automation tool that moves information around or the hosting the tool runs on.
• Upkeep. Providers update and retire older models, and your own processes change, so someone needs to check that the tool still works and adjust it now and then.

Ask for an estimate of the monthly running cost at your real volume, and a way to see actual usage once it's live. Ask whose name the accounts that bill for it are in, and what happens to them if the work ends. For our own projects, we send a quote after a call about what you need, and nothing starts until you've agreed to it.

That depends on which version of a tool you use and how many services your information passes through. The large AI providers, including OpenAI, Anthropic, and Google, say they don't use data from their paid business plans or paid developer access to train their models by default. Free versions can work differently. Some use what you type to improve their models, and some allow people to review it, depending on the product and your settings.

Paid plans can still keep records of requests for a limited time to check for misuse. And the AI provider isn't the only one involved. The person who builds the tool, the automation service that moves information around, and wherever the tool is hosted can all handle your data too.

A few practical steps cover most of this.

• Ask for a list of every service your information passes through, which plan each one is on, and who has access.
• Give the people who need AI a business account with the right settings, so nobody pastes customer details into a free personal account because it's the tool they have.
• If you handle information with its own rules, like health or financial records, check that the plan's terms cover it before anything goes in.

These are worth pushing back on before you sign.

• It doesn't name a task. Words like transformation, strategy, or innovation appear with no specific job attached.
• The main deliverable is a document or a roadmap, with nothing running in your business at the end.
• There's no agreed way to measure whether it worked, or the measure is vague, like adoption or engagement.
• It says the tool will be right every time, or that nobody needs to check its work.
• It doesn't say where your data goes or which outside services will handle it.
• Monthly running costs are missing or left for later.
• It recommends AI for everything, including tasks a simple form or rule would handle.
• It asks for a long commitment before anything has been tested on your real data.
• It doesn't say who owns the tool, the accounts, and the setup when the work ends.

1. 01What specific task will this change, and how will we know it worked?
2. 02What happens when the tool gets something wrong, and who catches it?
3. 03Where does our data go, which services handle it, and who can see it?
4. 04What will it cost to run each month at our volume, on top of the cost to build it?
5. 05Who owns the tool and the accounts when the work is done?
6. 06What would you not use AI for in our business?
7. 07Can we start with one task and decide on more after we've seen results?

Expect the tool to be wrong sometimes. Judge it by whether someone catches the mistakes before they cause a problem, and whether it saves more time than it takes to check.

Our AI consulting page explains how we approach this work, and the FAQ covers how projects are priced. If you have a task in mind, tell us about it.