Most businesses we talk to have already started. Someone bought a ChatGPT subscription, someone else prefers Claude, and both get used for drafting emails and rewriting proposals. That’s a reasonable place to begin. The frustration shows up a month later, when people realise the assistant knows a great deal about the world and nothing whatsoever about the business paying for it.
It can’t tell you which clients haven’t been invoiced. It can’t find the clause in the contract you signed in 2023. Ask it what happened on a job last month and it’ll happily invent something plausible.
That gap is closeable, and it’s usually cheaper to close than people expect.
What the subscription actually gives you
A chat subscription is a very capable model with no memory of your operations. It reads what you paste into it and forgets the rest. That’s why the value plateaus so quickly: every useful answer needs someone to first go and find the source material by hand, which is the part that was slow to begin with.
Both ChatGPT and Claude now let you attach files, build reusable project spaces and connect a few common tools. That covers real ground and it’s the first thing we check, because there’s no sense engineering a solution to a problem a setting solves. What it doesn’t cover is anything living in a database, a job system, a finance platform or a bespoke internal tool. For that, something has to fetch the record and hand it to the model.
Three ways to connect it to your data
The right route depends on where your answers live and how sensitive they are.
- Configure the workspace properly. Business or enterprise plan, admin controls, no-training terms, shared project spaces, document connectors, and a short policy staff will actually follow. Cheapest option, and for document-heavy teams it’s often enough on its own.
- Connect the systems through an API. We wire the model to your CRM, ERP, finance platform or job system so it can look up a live record before answering. The Model Context Protocol, which Anthropic published and OpenAI has since adopted, has made this considerably less painful than it was two years ago, though the work still sits in the data plumbing rather than the model.
- Build a private assistant. When the data is sensitive, the permissions are complicated, or the assistant needs to live inside your own software, we build it as a custom ChatGPT-style assistant over your business data with retrieval, citations and access rules designed in from the start.
We’ll usually recommend starting at the top of that list and moving down only when the questions demand it.
Asking your own data a question
The thing people actually want is to type a sentence and get a straight answer out of records that currently take twenty minutes to assemble.
That works well over documents. Contracts, policies, manuals, meeting minutes, proposals, email history, scanned PDFs. Retrieval finds the relevant passages, the model answers from those, and every answer links back to the file it came from so somebody can check it. Most of the value sitting in a business is in the unstructured material on the shared drive, and this is the first technology that’s been any good at reaching it.
Numbers need a different approach. A language model asked to total a column of figures will produce something confident and occasionally wrong, which is the worst combination available. The fix is to have the assistant write a query, run it against your actual database, and report the number the database returned. Same chat window, completely different reliability. Where the question is really a recurring metric rather than a one-off, a dashboard beats a chat interface and costs less to run.
One shortcut worth knowing. If the questions are about marketing spend and channel performance, don’t point a model at six ad platforms and hope. Consolidate them first, through a data hub like Funnel.io into BigQuery or Power BI, then let the assistant query one consistent set of numbers. We cover that side in dashboards and business intelligence.
Knowing which questions belong in which bucket is most of the skill, and it’s the part a vendor demo never shows you.
Inside the tools people already use
An assistant in its own browser tab gets used enthusiastically for about ten days. The ones that stick are the ones that turn up where the work already happens: a Teams or Slack channel, a button inside the CRM, a summary that appears on the job record without anyone asking for it.
That’s a plain systems integration job. The AI part is often the smallest piece of it. Getting the assistant into the workflow, with the right permissions and a sensible fallback when it doesn’t know, is what separates a tool people rely on from a subscription line item nobody can defend at renewal.
Where your data goes
The contracts are better than most people assume. Business and enterprise agreements from both OpenAI and Anthropic exclude your content from model training, and we’ll go through the specific terms with you rather than asking you to take our word for it. If your obligations rule out sending data offshore at all, the whole pipeline can run on open models inside your own environment. We cover the trade-offs in what private AI actually means and on-prem versus cloud AI.
The bigger risk in most businesses isn’t the vendor. It’s the personal accounts nobody has counted, quietly receiving client documents on a free plan. Sorting the accounts out and writing a one-page AI policy usually costs nothing and removes more exposure than any technical control.
Don’t marry a model
Providers leapfrog each other constantly, prices drop without warning, and access can be pulled for reasons that have nothing to do with you. We build with the provider call kept separate from everything else, so switching from Claude to ChatGPT, or to an open model you host yourself, is a configuration change rather than a rebuild. If you want the argument in full, we’ve written about lock-in and open models and how to choose a model for your business.
A good first project
Pick one question your team asks constantly and answers slowly. Where’s the current version of this document. What did we quote this client last time. Which of these jobs is waiting on a supplier. What does our policy say about this situation.
Connect the sources that answer that one question, get it in front of the people who ask it, and see whether they keep using it after the novelty fades. That tells you far more than a strategy workshop, and if the answer is no, you’ve spent a fortnight rather than a year finding out.
Bring us the questions your team keeps chasing, the systems the answers are buried in, and the rules about who’s allowed to see what. We’ll tell you which route fits, including when the honest answer is that a setting change and better habits will do the job.