How to use ChatGPT, Claude or Copilot
as your company brain.
Set a project up so it earns its keep, keep it current so it stays useful, spot the signs your team has outgrown it, and know what to put behind it when they have. Do this before you spend anything.
Brief
For a small, stable set of documents and a couple of decision makers, an assistant on its own is enough. Claude Projects and ChatGPT Projects both hold persistent instructions, attached files and project memory, and they are very good.
Set it up properly before you buy anything else. Write the instruction as a position rather than a persona, date everything you attach, delete more than you add, and share one project across the team. These four close a surprising amount of the gap.
The published guidance from both vendors names the boundary. Once you are past a handful of documents, or you need the same knowledge across different conversations and projects, they recommend an external knowledge base behind the assistant. We agree with them.
What breaks first is not intelligence. Nothing updates itself, nothing holds a position, and the knowledge stays in the account of whoever was most fluent with the tool. Liffey is the layer underneath, reachable from Slack, Claude, ChatGPT, Gemini, Copilot and Mistral, so the assistant stays the way in.
What each assistant does today.
Publicly documented capabilities as of August 2026, and this field moves monthly. Each of these is good at something specific, and knowing which one your team is in tells you where the ceiling is.
ClaudeProjects hold a persistent instruction plus attached files, with a large context window, project-scoped memory and automatic retrieval once the document set outgrows the window. Cowork adds agentic sessions on web and mobile, with project memory, scheduled tasks and work that continues while you are elsewhere.
Strongest atApplying an agreed method to new work, and multi-step work you would otherwise sit through.
ChatGPTEnterprise Projects give a team shared chats with common instructions, plus memory that carries facts between conversations. Connectors reach company systems, and admin controls and workspace policy sit over the top.
Strongest atGetting a whole team onto the same instructions quickly.
GeminiGrounds on Google Workspace: Drive, Gmail, Calendar and Chat. Gems let a team build custom assistants, and answers carry transparency citations that link back to the source paragraph or spreadsheet cell.
Strongest atTeams already inside Workspace, where the material lives in Drive and Gmail.
CopilotGrounds on Microsoft Graph and honours SharePoint permissions, so people see what they were already entitled to see. Copilot Studio builds task-specific agents, and Copilot Connectors pull in Salesforce, ServiceNow, Jira and Workday.
Strongest atOrganisations on the Microsoft tenancy, where permissions already model who sees what.
MistralLe Chat Enterprise runs with EU data residency, managed or self-hosted, with open-weight models available for air-gapped deployment. A large library of prebuilt MCP connectors reaches the usual company systems.
Strongest atEuropean teams where data residency is a requirement rather than a preference.
Liffey is reachable from every one of these through an MCP endpoint, alongside Slack, so the assistant your team already likes stays the way in. What matters is the quality of the context it reaches.
Setting one up so it earns its keep.
Most company projects get built in ten minutes and never revisited. An afternoon spent on the setup returns more than any other work you can do on a project, and it costs nothing. These steps follow the published guidance from the assistant makers and the practitioners who write about this.
- Create one shared project, and name an owner. In Claude, Projects then Create project. In ChatGPT, Projects then New project. One project the whole commercial team uses, with a named person who maintains it, is worth more than several excellent private ones, because knowledge in individual accounts leaves with individuals.
- Split behaviour from information. This is the step everything else depends on, and it is where most setups go wrong. Instructions define how the assistant should behave. Knowledge files carry what your company knows. Claude’s project instructions field holds roughly 8,000 characters, about 2,000 words, so treat it as a brief rather than a manual and put the facts in files.
- Write the instructions in the imperative, and say what to avoid. Direct commands get direct compliance: “Be concise” outperforms “I would prefer if you could try to be more concise.” Practitioners are consistent that telling an assistant which behaviours to drop works harder than describing ideal ones, because it is already trying to be helpful.
- State the company’s positions in the instructions. Who you serve, who you decline, what you charge and why, and what a good deal looks like. Most instructions describe a tone. Stating the decisions moves answers further than any other single change.
- Upload short knowledge files, one to three pages each. Concise, specific documents produce better answers than long ones, because the assistant extracts denser signal from tighter content. A twelve-page strategy deck is worth splitting into the four things it actually decided.
- Date and status-stamp every file in its first line. “Current as of March 2026” or “Superseded, kept for history”. Models weight what they can see and have no way to tell a current pricing sheet from last year’s unless you say so. This removes a whole class of confidently wrong answers.
- Delete more than you add. Superseded material competes with the current version for the assistant's attention. Remove a document the day it is replaced, and resist adding the replacement beside it.
- Write down what it is for. Tell people plainly which questions it answers well, so everyone knows the edge of what they are relying on.
You advise the commercial team at [company]. Use the knowledge files as the source of truth and say so when they do not cover the question.
Our positions: we serve [buyer] at [company size]. We decline [segment]. Pricing starts at [figure] and we hold it because [reason]. A good deal looks like [shape].
Be concise. Lead with the answer. Cite the file you used. Ask before assuming a number.
Adapt the shape, keep it under 2,000 words, and put everything else in files.
The upkeep nobody plans for.
A project degrades without anyone noticing. These are the habits that catch it, and each takes minutes rather than hours.
Book a recurring review
A repeating slot in one person's calendar to re-read what is attached and ask whether it is still true. Unglamorous, and it is the whole difference between a project that stays useful and one everybody stops opening.
CadenceMonthly, and after any decision that changes a position.
Keep a known-wrong test
Ask it about a position you superseded months ago. If it repeats the old answer with conviction, you have measured the gap on your own material rather than on somebody's marketing.
CadenceEvery review, same question, note whether the answer moved.
Watch who stopped using it
People rarely complain that an internal tool has got worse. They just go back to asking a colleague. That is your real quality signal, and it arrives long before anyone raises it.
CadenceAsk directly, once a quarter, of the people who used it most.
The vendors already tell you where the line is.
Published guidance for both Claude and ChatGPT Projects says the same thing: once you are past a handful of documents, or you need the same knowledge available across different conversations and projects, you want an external knowledge base behind the assistant.
The people who build the assistants drew that boundary, and we agree with them.
It is worth understanding why the line falls there rather than somewhere else. An assistant is designed to be excellent inside a session: hold a lot, reason well, produce something good. A company's commercial knowledge has to be right across every session, every person and every quarter. Those are different requirements, and the second is a system rather than a setting.
You will know you have reached it when re-uploading the latest version has stopped happening, when two people get different answers to the same question, or when somebody acts on something that was true last spring.
What a project cannot carry for you.
None of these are criticisms of the products. They are consequences of what a project is for, and no amount of prompt work removes them.
It waits to be updated
Somebody re-uploads the file when it changes. That happens for a month, then sometimes, then it stops. The assistant carries on answering with total confidence from the version it has.
What to do about itName an owner and a review date, or accept it will drift.
It holds files
It can tell you what a document says. It cannot tell you that a pattern across six accounts contradicts the buyer definition you agreed in March, because it was never told you had one and nothing is watching for the contradiction.
What to do about itState the positions in the instruction, and revisit them deliberately.
It stays where it was put
Knowledge accumulates inside individual projects and individual accounts. The best version usually belongs to whoever is most fluent with the tool, and it leaves with them when they do.
What to do about itOne shared project, maintained by a named person, from the start.
Pointing an assistant at your drives is not curation.
The instinctive next step is access: connect the assistant to the drive, the CRM and the shared folders, and let it read everything. It helps, and it also brings back everything at once, whether or not it is still true.
What comes back is the closest match across a decade of files, weighted by nothing in particular. The pricing deck from two strategies ago sits beside the current one. The proposal somebody drafted and abandoned reads exactly like the one that won. Scaling that across a team means every person needs to already know which answer to trust, which is precisely the knowledge you were trying to get out of their heads.
Settle these before you widen access:
- Who can reach what, and whether that mirrors the permissions in the original tool
- What happens to that reach when somebody changes team or leaves
- How long personal data in call transcripts and email threads is kept, and how it is removed on request
- Where content is processed and stored, which in Europe means a named region
- Whether anything you attach is used to train a model, which differs by plan and by vendor
That work exists whichever route you take, and settling it deliberately beats meeting it in a security review.
Curation is the part access alone will not give you. In a Liffey brain every source is chosen by a person: you and your architect nominate what goes in, and each source records who added it and how it should be treated, so an answer traces back to material somebody put there on purpose.
When to add a layer underneath.
This is an addition rather than a swap. The teams that get the most out of a layer are usually the ones already deep into Claude or ChatGPT, because they have felt the ceiling and can describe exactly where it is.
The assistant on its own is enough when
- Your document set is small, and it is fairly stable
- The knowledge is reference material rather than a position that has to be agreed
- One or two people make the commercial decisions anyway
- The people using it are fluent enough to spot a stale answer
- You are early, and the strategy is still moving too fast to write down
- Nobody outside the founding team needs the same answer yet
Put something behind it when
- Re-uploading the latest version has already stopped happening
- More than a handful of people need to reach the same answer
- The knowledge has to survive somebody leaving
- What matters lives in calls, deals and threads rather than in files
- Nobody has the job of noticing when the strategy and the evidence disagree
- You need to say who can reach what, and show it
Your company’s commercial brain
and a senior operator to maintain it.
A senior commercial architect, the method they work from, and the brain that holds what your company knows about winning.
The architect
A senior commercial operator, twenty years from scale-up to PLC, working alongside your team on the strategy itself. They judge what the evidence means before anyone acts on it, and that judgement is a person's job.
- From 20 hours a month
- Has carried a number, set pricing, built a motion
- How an engagement runs
The method
A structure that says how the parts of a commercial strategy relate: the value you deliver, the buyers it is worth most to, how you reach them, and what happens after the win. It is the full version of what a good project instruction aims at.
- Built from work with companies across Europe
- Your version of it is yours to export
- The method in full
The platform
It reads the systems you already run, holds what matters about how you win, and answers in plain English with the evidence attached. It updates itself, so nobody has to remember to re-upload anything.
- Ask from Slack, Claude, ChatGPT, Gemini, Copilot or Mistral
- An MCP endpoint, a versioned REST API and a CLI
- Calls, deals and threads, as well as documents
The clearest way to judge it is to watch one running. There is a recorded walkthrough, and a five-minute diagnostic that names the weakest part of your commercial picture without you talking to anybody.
Questions people ask.
Can I use ChatGPT, Claude or Copilot as a company brain?
For a small, stable set of documents and a couple of decision makers, yes, and Projects are a good way to do it. Set it up like this and you will get most of what is available:
- One shared project, with a named person who maintains it.
- An instruction that states positions, not a tone: who you serve, what you charge, why you win.
- Decisions attached rather than collateral, each one dated and status-stamped in its first line.
- A monthly review that removes what has been superseded rather than adding beside it.
- One known-wrong question you re-ask each review, to catch drift before your team does.
What are the limits of Claude Projects and ChatGPT Projects?
There are three practical limits. Nothing updates itself, so somebody has to re-upload a document when it changes. Nothing holds a position, so the project cannot tell you when new evidence contradicts a decision you made. And knowledge accumulates inside individual accounts and projects rather than in the business, so it leaves when people do.
How do we stop an AI assistant giving confidently wrong answers?
Date and status-stamp every file in its first line, remove superseded material rather than letting it compete, and keep the attached set small enough that a person could still read it. Most confidently wrong answers in a project trace back to two documents disagreeing with no way to tell which one won.
Should we connect our AI assistant to our drive and CRM?
It helps, and settle governance before you widen access: whether permissions mirror the source tools, what happens when somebody leaves, how long personal data is kept, where content is processed, and whether anything is used for training. Broad access without those answers is the version that surfaces during a security review.
How do we keep an AI assistant setup EU-resident?
With an assistant, residency is decided by every system in the chain rather than by the assistant alone. The vendor’s own hosting region is one part of it. The others are each tool you connect, wherever that tool processes and stores what it holds, plus wherever the model itself runs. A single connector outside the region changes the answer for the whole set, so the honest way to establish this is to list the systems and check each one.
Liffey was built EU-resident from the ground up: database, application compute, AI inference and embeddings all run in Frankfurt, adhering to ISO 27001:2022 and GDPR. Named sub-processors and their locations are published on the sub-processors page, and the architecture is set out on the security page.
What happens to personal data in the call transcripts and emails we give an assistant?
This arrives whether you plan for it or not, so decide it before you widen access. The question to answer for any setup is where identifiers are removed, and how completely a deletion request can be honoured once material has been copied into passages and indexes.
In Liffey every workspace can enable redaction at ingestion, so identifiers are removed before storage, indexing or AI processing. Liffey acts as a processor under an Article 28 agreement, and the terms are in the data processing agreement.
How do we stop last year’s information outweighing this quarter’s in an AI assistant?
Left to itself, an assistant weighs whatever it can see, so a pricing sheet from two strategies ago carries the same authority as the current one. In a project you manage that by hand: date every file, and remove what has been superseded on the day it is replaced.
Liffey builds recency in. Evidence accumulates against the pattern it belongs to, weighted so recent evidence leads while older evidence stays in the count, and it retires on a defined schedule. So a pattern can build across quarters while the current position still wins a close call.
Do ChatGPT, Claude and Copilot train on our company data?
The answer follows the plan rather than the brand. Business and enterprise tiers of the major assistants exclude customer content from model training by default, and consumer tiers vary, with some defaulting the other way. Two things are worth having in writing: what the tier you are actually on does, and what is retained once an answer is returned, which is a separate setting from training.
Liffey holds enterprise terms with Anthropic and AWS that exclude your inputs from model training, and the AI layer runs with zero data retention, so prompts and responses are processed in memory.
How do AI assistant permissions work once we connect our drive?
Where the connector supports it, the assistant respects the permissions of the source tool, so the live question becomes what those permissions currently say. Most drives carry years of broad sharing that stays quiet while search is weak, and a capable answer engine surfaces it immediately. Two things to do before you widen access: check what “anyone with the link” covers today, and agree what a leaver triggers across both the assistant and anything indexed from it.
In Liffey you choose the scope at connect time, down to the channel, the repository path and the part of a drive, and widening it is a separate deliberate act. Membership is invite-only, and the audit log is append-only, recording the actor, the time, and the before and after state.
Which AI assistant should we choose as a company brain?
Route it by where your material already lives. A Microsoft estate points to Copilot, for Graph and SharePoint permissions. Google Workspace points to Gemini. Long documents and careful reasoning over an attached set point to Claude. The broadest tooling around the assistant points to ChatGPT. European residency or self-hosting as a firm requirement points to Mistral.
Any of them will do this job. At this stage the difference that matters is how little friction there is in keeping your material in front of it, which is decided by which tools your team already has open all day.
Do we have to stop using ChatGPT, Claude or Copilot if we adopt Liffey?
No, and we would rather you did not. Liffey is reachable from Slack, Claude, ChatGPT, Gemini, Copilot and Mistral through an MCP endpoint, so people ask in the tool they already have open and get an answer grounded in your own agreed context.
Can Claude Cowork act as a company brain?
Cowork is a genuine step: agentic sessions with project memory, scheduled tasks and work that continues while you are elsewhere. It makes the doing much stronger. It starts from whatever context you give it, so the stronger that layer is, the more you get out of Cowork.
How do we know when we have outgrown an AI assistant as our company brain?
Three signals, and one is usually enough. Re-uploading the latest version has stopped happening. Two people ask the same question and act on different answers. Or somebody makes a decision from something that was true last spring and nobody caught it.
What does Liffey add that a very good prompt does not?
A structure that holds the company's positions and how they depend on each other, updated from your own systems rather than by somebody remembering to re-upload, and a senior commercial architect who judges what new evidence means before the team acts on it. A prompt is a set of instructions. Liffey is the agreed knowledge those instructions work from.
How much does Liffey cost compared to our assistant licences?
There is a free tier. From €2,995 a month, a senior commercial architect operates fractionally in your business, evolving and being accountable for the strategy that wins and grows your customers. It is a separate line from your assistant subscriptions, and the architect’s time is included. Full detail is on the pricing page.
“We’ve grown by nearly 50% over the past year. In that time, Liffey has helped us better articulate our unique value proposition, evolve our pricing model, reset our go-to-market strategy and build scalable commercial systems. All of which has enabled us to make better decisions on a day-to-day basis and grow with more confidence.”
Nick Comer, Founder & CEO
Twenty minutes on your commercial strategy.
Where the strategy is thin, what belongs in a commercial brain, and whether your assistant is already covering it. If it is, we will say so.