Everyone says we should be using AI and nobody can tell me where it would actually save us money.
Two questions, in order: which processes are worth automating, and what happens when it gets one wrong. We answer both, then build the smallest thing that works and prove it. Most of the money lost on AI in businesses this size is not lost on the technology. It is lost on tools bought by individual departments, paid for on cards, pointed at customer data nobody mapped, and impossible to unwind by the time anybody asks a question about them.
Dinmore Bell governs AI adoption for founder-led businesses: identifying where a model genuinely helps, procuring it properly, and establishing the data-protection position, intellectual property in inputs and outputs, human oversight and the record of decisions. Model development is delivered by specialist partners under its procurement and governance.
How we own it
Find the work
Volume, repetition and clear rules. Contract review, first-line customer response, document extraction, reporting. Most businesses have three obvious candidates. The way to find them is to look for the work that is already being done by a person following a procedure, where the input arrives in a consistent shape and the output is checked by somebody anyway. That last condition is the important one: a process with a review step already in it can absorb a model that is right most of the time. A process with no review step cannot, and putting one in front of a customer is how a business discovers the difference in public.
Govern it
Data flows, contractual permission to use client data, human review points, audit trail, and the policy your enterprise customers will ask for. The first of those is the one that is usually skipped. Before a model sees anything, somebody has to be able to say what data goes in, where it is processed, who else touches it, how long it is kept, and on what basis you were entitled to use it for this rather than for the thing the customer originally gave it to you for. That is a short document when it is written at the start and an archaeology project when it is written eighteen months later.
Prove it small
One process, measured against the manual baseline, before anything is rolled out or renamed as a transformation programme. The baseline is the part people skip, and without it there is no honest way to say whether the thing worked. Measure the current cost in time and error rate, run the model against the same work for a fixed period with the human review still in place, and compare. A pilot that cannot show the comparison is not a pilot, it is a purchase with a friendlier name.
Buy it properly
AI procurement fails on the same three clauses every time. Training rights: does the vendor use your inputs to improve its models, and can you switch that off in the contract rather than in a settings page. Location and sub-processors: where does the processing actually happen, and does the vendor have to tell you before that changes. Exit: what happens to your data and your outputs at the end, and how do you get them back in a usable form. Add the commitments the vendor makes about accuracy and availability, and note that a free tier usually makes none. We negotiate those before the tool is embedded in a process, because afterwards the leverage has gone.
Who owns what the model produces
Intellectual property in AI output is genuinely unsettled, and a business should plan on the assumption that some of what a model produces may not attract the protection its own work would. That has practical consequences. It affects what you promise customers about the originality of deliverables, what you can sensibly claim to own in a diligence process, and how you handle work product created by a contractor using their own tools. Our starting position is contractual rather than doctrinal: say clearly in the customer agreement and in the contractor agreement who owns what and who bears the risk if a third party complains.
The policy your customers will ask for
Enterprise buyers now ask whether you use AI, where, on what data, and with what human oversight, and increasingly they ask it in the same security questionnaire as everything else. An honest two-page answer, with a named owner and a register of the tools actually in use, closes that question. A policy written to sound reassuring and contradicted by what the engineering team is doing closes nothing, because the follow-up question is always evidence. We write the policy against what the business actually does, and keep the register current as tools change.
Oversight that is real rather than stated
Human oversight means a named person who can see what the model decided, understands the decision well enough to disagree with it, and has the authority and the time to override it. Each of those can fail separately. A reviewer with no visibility into the reasoning is a rubber stamp; a reviewer with fifty decisions an hour to approve is a rubber stamp with better intentions. We design the review step as part of the process rather than as a line in a policy, and we record the overrides, because the pattern in what gets overridden is the most useful information the pilot produces.
What this costs, and where the specialists come in
A scoped engagement to find the work and set the governance, then a fixed monthly retainer to run it as the estate grows. Model development, data engineering and the build itself are delivered by specialist partners under our procurement and governance, so the business has one party accountable for the commercial and legal position and a named team accountable for the technology. Regulated advice on tax treatment, and any question requiring a regulated legal opinion, comes from instructed specialists whom Dinmore Bell prepares, coordinates and controls.
Where specialists come in
Model development and engineering are delivered by specialist partners under our procurement and governance. Data-protection impact assessments involving regulated processing are reviewed by instructed specialists.
Related results
All results →Sectors we run this for
Related insights
All insights →Where nine per cent of your revenue goes
Contract value rarely leaks in a single expensive event. It leaks quietly, through renewal dates nobody owns.
What a fully loaded General Counsel actually costs
The comparison is not our fee against an hourly rate. It is our fee against the hire you would otherwise make.
- Where do most businesses actually save money with AI?
- In work that is high volume, repetitive, rule-governed and already reviewed by a person: document extraction, first-line customer response, contract triage, routine reporting. The saving comes from shortening a process that already exists, not from inventing a new one.
- Can we put customer data into a commercial AI tool?
- It depends on what you told the customer you would use their data for, what your contract with them says, and what the vendor does with inputs. All three need checking before the first upload, and the answer is often yes with a configuration change and a contract amendment.
- Do we need an AI policy?
- If you sell to enterprise customers, you will be asked for one, usually inside a security questionnaire. More usefully, the exercise of writing an honest one is how most businesses first discover which tools are already in use and on what data.
- Who owns the output a model produces for us?
- The legal position is unsettled and varies by jurisdiction and by how much human input shaped the result. We deal with it contractually: say expressly in the customer and contractor agreements who owns what and who carries the risk of a third-party complaint.
- What does meaningful human oversight look like?
- A named person who can see the basis of the decision, understands it well enough to disagree, and has both the authority and the capacity to override it. Overrides should be recorded, because what gets overridden tells you more about the model than the accuracy figure does.
- How long does a pilot take?
- Long enough to produce a comparison against the manual baseline, which usually means four to eight weeks on a single process. Anything shorter produces an anecdote rather than evidence, and anything longer is generally a rollout that has not been called one.
- Do you build the models yourselves?
- No. Model development and data engineering are delivered by specialist partners, procured and governed by Dinmore Bell. We are accountable for the commercial terms, the data position and the oversight; they are accountable for the build.
- We already have tools in use that nobody approved. What now?
- Start with the register: what is in use, by whom, on what data, and under whose contract. In most businesses that exercise takes a fortnight and produces two or three decisions that need making immediately, and it is considerably cheaper than finding out during a customer audit.