Your company's ChatGPT is not an AI strategy
Individual licenses create personal gain and organizational loss. What a company forfeits when the knowledge produced with AI dies inside the chat window.
*First article in our series on enterprise AI, legacy integration and information governance.*
Executive summary
Most companies that say "we already use AI" bought licenses. That solves individual productivity and builds no corporate asset: the reasoning, the context and the decisions generated across thousands of conversations sit in personal accounts, outside systems of record, outside audit, and out of reach of every other team. The result is a company that pays for AI every month and accumulates nothing. Strategy begins when AI stops being a subscription and becomes an architectural layer — connected to legacy systems, with access control, an audit trail, and reuse of what was produced.
What a license solves, and what it does not
Tools like ChatGPT, Claude and Gemini are excellent at what they promise: speeding up one person on one task. An analyst drafts in 10 minutes what used to take 2 hours. A manager summarizes an 80-page contract. An engineer debugs an obscure log.
None of that is small. But notice what happens next: the analyst pastes the text into Word, the manager drops the summary into an email, the engineer closes the tab. The intellectual work behind the result — the context supplied, the discarded attempts, the judgment applied, the model version used — disappears.
If the same question resurfaces three months later in another department, it will be answered from scratch. The company paid twice for the same reasoning and has no way of knowing it.
| Layer | Individual license | Enterprise platform |
|---|---|---|
| Gain | Productivity per person | Productivity + reusable asset |
| Context | Pasted manually by the user | Retrieved from company systems |
| Record | Vendor's personal account | Repository under company control |
| Access | Depends on user discipline | Inherited from corporate permissions |
| Audit | Effectively nonexistent | Input, output, approver and model version |
| Scale | Linear with license count | Compounding: each use improves the next |
Shadow AI: the diagnosis almost nobody ran
Before debating platforms, measure what is already happening. In practice, nearly every mid-size and large company already has AI usage outside IT control — personal accounts, browser extensions, and features embedded in SaaS bought department by department.
Three questions that tend to produce uncomfortable answers:
- 1.How many distinct AI tools are used for work today, including the ones nobody approved?
- 2.What kinds of data have already left the company inside a prompt — contracts, customer records, source code, financial projections?
- 3.If a client or a regulator asks for the trail behind an AI-assisted decision made in the last 12 months, can the company produce it?
If the third question has no answer, the problem is not tooling. It is information architecture — the subject of Where what your employees type into AI ends up.
The four invisible costs
Evaporated knowledge. The most expensive asset a company produces with AI is context: what was tried, what worked, why. In isolated usage, that context has a half-life of one session.
Silent rework. With no shared repository, different teams solve the same problem in parallel, at uneven quality, unaware of each other.
Unpriced risk. Sensitive data pasted into a chat window becomes an exposure nobody inventoried. The cost arrives all at once, in an incident, not in monthly installments.
Vendor dependency. If all generated value lives inside one vendor's account, the company has no exit strategy — it has a mandatory renewal.
What a real AI strategy looks like
The right question is not "which model should we use?" but "where will the generated knowledge live, who can access it, and how does it flow back into the business?". That implies four decisions:
- 1.Context source. AI must read the systems the company already runs — ERP, CRM, PLM, document repositories — honoring existing permissions. That is the subject of *Your ERP is 20 years old and AI needs to talk to it*.
- 2.Ownership of the record. Prompts, responses, cited sources and approvals live in infrastructure the company controls, not in personal accounts.
- 3.Governance proportional to risk. Low-risk use cases take a fast lane; decisions affecting customers, money or people go through recorded human review.
- 4.Portability. The model is a replaceable component. Context, system prompts, evaluations and history must survive a vendor change.
What to do on Monday
- Request an AI usage inventory by department, including unapproved tools. Two-week deadline, no penalties — the goal is visibility, not discipline.
- Pick one high-volume, high-context process (sales proposals, contract review, technical support) to become the first case with a corporate record.
- Define a simple data classification rule now: what may never enter an external tool.
- Put the audit-trail question on the next executive committee agenda, before a customer puts it there for you.
Conclusion
Buying licenses is a cost decision. Building the layer where generated knowledge is preserved, audited and reused is a strategic one — and it is what separates the company that merely gets faster from the company that gets smarter every quarter.
Further reading
Engineering track:
- Reference architecture for enterprise AI: layers, contracts, and perimeter — the technical deep dive on this topic.