A prompt engineering consulting service writes, tests and hands over the instructions an AI model follows in one of your workflows. Pay for it when a specific workflow already runs on a model, you can count how often it goes wrong, and someone on your side will own the prompts afterwards. Skip it when the real risk is what the software is allowed to do, because a prompt cannot enforce a limit.
I should say where I stand. I offer agent builds myself, and prompt work is part of each build. I don't sell a separate prompting service, and I do have an interest in you buying an agent. Weigh what follows with that in mind.
What an AI prompt consultant should deliver
Service pages that rank for this search tend to list custom prompts, prompt optimisation, strategy advice, testing and ongoing monitoring. Those are activities. Ask what you will hold at the end. I would expect six things.
- Prompt files in your repository. One file per workflow, in plain text, with a comment saying what each rule is for. A prompt that exists only in a vendor console or a slide deck is hard for your team to change.
- A test set. Real cases from your workflow, anonymised where needed, each with the answer a competent colleague would give. You can rerun it after every later change, so it keeps its value longer than any single prompt.
- A score before and after. The old prompt and the new one run against the same test set. Without the first number, the second one proves nothing.
- Hostile cases. Inputs with instructions hidden in them, such as a forwarded thread or an email signature that tells the model to ignore its rules. They run again whenever the prompt or the model changes.
- A written list of what stays out of the prompt. A consultant who has looked at your workflow can name the rules that need to sit in code. Treat an empty list as a warning sign.
- A way to change the prompts. Edits go through review with a diff and a revert, and the log records which prompt version produced which output.
A workshop on writing better prompts for chat tools is a different purchase. It can help a team that uses those tools all day, but buy it as training and judge it as training.
When paying for it makes sense
A live workflow fails in ways you can count. A model sorts incoming mail, extracts fields from PDFs or drafts replies, and your team corrects the same kinds of mistake every week. You already have the raw material for a test set and a number to improve.
The mistakes are about judgement and format. Tone, classification, the structure of a reply, the point at which a case goes to a person. These behaviours live in the prompt, so changing the wording changes them.
A model change is coming. A rule that held under one model can slip under its successor. Someone has to rerun the cases and fix what moved. If your team has never done that, paying once to set up the routine is reasonable.
Someone on your team will own the result. That person reads the test results and edits the files. Without one, nobody notices when the prompts stop matching the work.
When it doesn't
The software can do something nobody can undo. Sending external mail, posting to a ledger, deleting a record. A good model follows an instruction almost always, and almost always is the wrong standard for those actions. The limit belongs in permissions and code, which is the argument in a prompt is a request, a guardrail is a wall.
Nobody can say what correct looks like. If two experienced colleagues give different answers to the same case, no prompt can satisfy both. Settle the rule first.
The model is missing information. If the order number isn't in the email and the model can't reach the order system, better wording produces a more confident guess. That is an integration job.
A large share of cases needs a person's judgement. Prompts won't move that share by much. The same fact decides whether an agent is worth building at all.
You have no workflow yet. You would be tuning instructions for work that may be the wrong place to start. What to automate first comes before any of this.
Where prompt work stops
An AI system that does real work has three layers, and prompt consulting mainly touches the first.
| Layer | What it does | What it can't do |
|---|---|---|
| Prompt | Shapes how the model reads a case and what it writes | Guarantee anything, because a model's output varies from run to run |
| Evaluation | Measures behaviour on a fixed test set and shows when a change made things worse | Stop a bad action in production |
| Guardrails | Limit what the software can do, through code and permissions | Improve the quality of an answer |
Here is an invented example, with no client behind it. A company receives 1,500 support mails a month, and a model sorts them into queues. On a test set of 200 past mails, the old prompt sorts 170 correctly and a consultant's rewrite sorts 188. Wrong answers fall from 30 to 12. If the test set represents the real inbox, misfiled mail falls from 225 to 90 a month.
That result is worth paying for if each misfiled mail costs your team real time. It also leaves about 90 wrong answers a month. For sorting, a person moves the mail to the right queue and nothing is lost. Give the same model the job of issuing refunds and that error rate becomes unacceptable, however good the wording. You would then keep the payment permission away from the model and have a person release each refund.
The test set tells you how many errors remain. The guardrail decides what those errors can do. A proposal that covers prompts and leaves out evaluation gives you new wording and no way to tell whether it helped.
A test set only covers past cases. A shadow run tests the prompt on live input while the software changes nothing.
What to ask for before you sign
- Which workflow and which number will this improve? A good answer names one workflow and one measure. A weak answer is "better AI output across the business".
- Who builds the test set, and who keeps it? A good answer is your team and the consultant together, stored with your files. A weak answer is a proprietary framework you can't take with you.
- What is the score today? A good answer is a measurement before any rewriting starts. A weak answer is an estimate from a demo.
- Where do the prompts live when you leave? A good answer is your repository, in files your team can read. A weak answer is an account only the consultant can open.
- What happens at the next model update? A good answer is a rerun of the test set by someone you can name. A weak answer is that good prompts work on any model.
- What would you refuse to put in a prompt? A good answer names the risky actions in your workflow and says where those limits go. A weak answer is "we can add a strong instruction".
- Will real cases leave our systems during this work? A test set of customer mail contains personal data, so the consultant needs the same scrutiny as any AI vendor. These twelve privacy questions apply.
How this relates to building an agent
A prompt is one file inside an agent. A build also needs the connections to your systems, the permissions, the review queue, the log of every action and the handover. If your software only answers questions, prompt work alone may be enough, and how an AI agent differs from a chatbot helps you tell which you have. Software that acts needs the rest.
In an agent build from me, each agent gets a prompt file for the rules of its workflow, plus a shared file for the rules every agent follows. It runs in shadow mode and has to clear a pass mark we agree up front. A file of prompt injection cases runs again whenever the prompt, the code or the model changes. The prompts end up in your repository, where your team edits them through pull requests.
What drives the cost of a custom build is in how much an AI agent costs to build.
Take twenty cases your workflow got wrong, anonymised, to a thirty-minute scoping call. We sort them into the ones that need new wording, the ones that need data the model can't reach and the ones that need a rule your team hasn't settled. If new wording is enough, I will tell you so.