AI services
AI Operations
We put AI into named day-to-day workflows, with the data path agreed in writing, and leave the systems running in the firm's own accounts when we go.
The issue
The problem shows up differently depending on the firm. These are the questions we hear, in the words clients actually use, and the answer we would give in the room.
- Legal
We have seen the demo. Nobody will sign off the tool. What would we actually run on Tuesday?
- Accounting and finance
Which of our hours can a model touch without client files leaving the building?
- Medical and scientific
We will not have a chatbot answering patients. What can AI safely do inside the clinic?
- Travel and hospitality
Guest and payment data is not a playground. Can any of this still save time?
We put AI into work the firm already does. We take a named workflow, agree where the information goes in writing, and build only what the partners can defend. The systems stay running when we leave, in the firm's accounts, with documentation a future IT lead can read. This is AI consulting in the sense buyers mean it, and delivery in the sense we mean it.
A chat box that answers clients' legal, tax or clinical questions creates a data path you cannot defend, and we will not build one. A triage assistant that answers the basic questions a visitor would otherwise put to a fee earner, declines the specific question, and hands the person to you is a different workflow, and we run one on this site. Some workflows are declined. That is part of the service. If you do not yet have a written position per workflow, we start there, usually as the AI governance engagement, because building first and asking permission afterwards is how firms end up with a tool they cannot defend.
How this looks in your firm
In a law firm, the demo went well and nobody will sign off the tool, so the useful question is what you would actually run next week, with client files in the building.
In an accountancy, the hours that look automatable often sit next to data that cannot leave.
In a medical or scientific practice, a chatbot that answers patient questions is refused. What remains is a triage assistant that only says who the clinic treats and how to book, and named internal workflows with a data path you can show.
In travel and hospitality, guest and payment data is not a playground, and operational volume still needs that same written decision.
Legal, accounting and medical feel the obligation first. Travel often has the volume. Both still need a clear yes or no before anyone builds.
Access credentials and vendor contracts sit in the firm's name, not ours. Handover includes who to call when something breaks, which should be your people first and us second. If the honest answer is that a workflow should stay manual, we write that down and do not build it.
What you get
These are the concrete things the engagement delivers, not only a report and an invoice.
A short list of candidate workflows, with a written go or no-go for each.
Implementation of those that clear, handed over running, with the vendor relationship in the firm's name.
For each live workflow: where the prompt goes, whether it is retained, and how you would show the trail twelve months later.
Training for the people who will operate it, and a written reason for every workflow we decline.
MeasurementHow we will know it is done
The project is done when the cleared workflows are in use by the named team, the declined ones are documented, and a partner can explain both lists without us in the room.
Time saved is a useful internal number and a poor external claim, so we only report it if the firm is already measuring that work.
The test we will always report is whether the data path still matches the written position thirty days after handover.
If it does not, the build is not finished.
Workflows we decline stay on the list, with the reason, because the refusals are what give the approvals their weight.
Who this is not for
If you want a chatbot that answers your clients' legal, tax or clinical questions on the website, or nobody in the firm can say which work the tool should be allowed to touch, we are not the right people. We build tools that run real work every day, not demonstrations that never go live.
An example of this work
Wanted AI in the practice without breaching a single obligation
Multi-office accountancy group, eight offices. Anonymised.
The partners were not asking what AI could do. They were asking what they could defend to a regulator and to a client, which is a different question and a harder one.
Questions
Will you put a chatbot on our website?
Not one that gives advice. A chat box that answers clients' legal, tax or clinical questions creates a data path you cannot defend and a liability you do not want. A triage assistant is different work, and we run one on this site: it answers the basic questions a visitor would otherwise put to a fee earner (what the firm does, who it acts for, what it costs, how to start), declines the specific question, and hands the person to you. It is built like any other workflow, with a written position on where the text goes and what is kept. If that is what you mean by a chatbot, it is in scope. If you mean a bot that advises, it is not.
What do you hand over?
A running system for the workflows that were cleared, documentation of the data path for each, and a written position the partnership can show a regulator or a client if they ask. Access credentials and vendor contracts sit in the firm's name, not ours. We do not leave you inside a black box you cannot operate. If the honest answer is that a workflow should stay manual, we write that down and do not build it. Handover includes who to call when something breaks, which should be your people first and us second.
How does this sit next to governance?
Operations builds what governance has cleared. If you do not yet have a written position per workflow, we start there, usually as the AI governance engagement, because building first and asking permission afterwards is how firms end up with a tool they cannot defend. Where a small operations piece is obviously safe, we will say so and not force a longer review. The two services are separate on purpose: one decides, the other delivers.