We've been working with a lot of organizations about how to roll out AI across large teams. Most are dealing with some version of the same situation:
- Employees want access to whichever AI tool works best for them or that they have a personal affinity towards.
- Security and IT need control, consistency, and visibility.
- Procurement wants leverage and predictable costs.
- Communications wants accuracy, brand consistency, and responsible use.
- Leadership wants adoption and measurable returns.
- Meanwhile, the capabilities, costs, and policies of the models keep changing (daily may be too far of a stretch… but feels like it).
All of those positions are reasonable.
Employees are already finding useful ways to work with AI, and they do not want to lose access to tools that help them. IT cannot reasonably support every new product someone discovers. Procurement cannot negotiate and manage dozens of separate agreements. Communications cannot have every team working from different facts, different messaging, or a loose interpretation of the brand.
So the natural response is to choose one provider and make it the company standard.
That may be the simplest answer administratively. I am just not convinced it is a good idea to make that choice difficult to undo, especially right now.
The market is changing too quickly
The relative quality of the models changes constantly. A model that is clearly better at one kind of work today may be overtaken within a few months. Some are better at writing. Some are better at research or analysis. Some handle long documents, images, or technical information better than others. Prices and usage limits change just as quickly.
Large organizations also have more varied needs than most vendor demonstrations acknowledge. The best tool for a communications team may not be the best one for legal, research, operations, or technology. Even within communications, the model that is good at creating an initial draft may not be the one you would choose to analyze a large body of source material.
I am not suggesting that organizations avoid enterprise agreements. Those agreements provide much better security, privacy terms, support, and pricing than employees using personal accounts. And for a lot of organizations we work with, I do advocate for using a standardized tool across a team or multiple teams when the use case fits it.
I would, however, be careful about any agreement or implementation that assumes one model will meet every need for the next several years.
The question I would ask is: if we want to use a different model a year and a half from now, how difficult will that be?
Will we be changing a setting, or rebuilding the entire system? What are the costs associated with changing a model or system, versus the costs of being locked in to one for too long in a sector that is changing so rapidly?
Pay attention to why employees prefer certain tools
Different people inside an organization will have their favorites. It is tempting to see that as another form of shadow IT that needs to be shut down.
Sometimes it is. People should not be placing sensitive information into unapproved systems.
But their preferences can also tell you something useful.
If a team keeps returning to a particular tool, ask what it does better for them. Maybe the approved tool is poor at working with long documents. Maybe it loses too much nuance when editing. Maybe employees cannot access it easily. Maybe one team has developed a genuinely good workflow that nobody responsible for the global rollout has seen. Or maybe if someone reads “and honestly, this changes everything…” one more time they may burn the office down.
You do not need to approve every tool people request. I just encourage the question to understand what is driving the request before deciding that one company-wide product has solved the problem. There might be some hidden insight there.
Decide what actually needs to be consistent
For most large organizations, the most universal list includes:
- Who can access company information
- Which information they can access
- Which sources are current and approved
- How confidential and personal information is handled
- When human review is required
- Which types of work should not use AI
- How employees report a bad or questionable result
Since we work so closely with comms and marketing teams, I’d add a few more baseline questions, especially if AI work will be public facing:
- Are people working from current information?
- Are they using the same approved messages and positions?
- Is generic off-brand creation causing reworks, upward delegation, or brand or reputational risk?
- Do teams know what flexibility they have to adapt?
- What governance in brand is considered ‘close enough’ versus ‘too far’?
- Is outdated information returning in new content?
All of those things should be consistent across an organization. They also belong to the organization, rather than to any particular AI provider.
So when I think about what should be consistent, I am thinking about the company context people are working from: the facts, policies, messaging, voice, examples, visual guidance, and review expectations. I am also thinking about who can access that information and how sensitive material is handled.
The model is one part of the system. It can be an important part, but it is also the part most likely to change.
Ideally, employees have a clear, approved place to work with AI, while the organization retains some choice over what is powering that work. That does not mean offering every available model. It could mean using one model for most work and maintaining one or two others for cases where they are demonstrably better or meet a particular security or other requirement.
For the employee, the experience can still be simple. The complexity does not need to be passed on to them.
Test the work you actually do
I would not select a company-wide AI tool based mainly on general benchmarks or product demonstrations.
Use your own material and your own tasks.
For a communications team, that might include:
- Summarizing a difficult internal document without dropping important qualifications
- Drafting from approved sources without inventing additional claims
- Adapting language for different audiences while preserving the meaning
- Comparing a new statement against previous company positions
- Identifying where a draft needs legal, policy, or regional review
- Working with the terminology your organization actually uses
- Recognizing when the available information is incomplete or contradictory
The results will tell you much more than a generic demonstration. We’ve spent an obscene number of hours scrutinizing and updating quality metrics and audit skills for AI output. It’s just such an important part of any lasting AI enablement.
Good testing and good auditing skills will also give you a way to revisit the decision later. If another model becomes available, you can run the same work through it and compare the results. You do not have to rely on employee preference, vendor claims, or whichever model is getting the most attention that month.
This testing does not need to become a huge internal program. A good set of representative tasks and skills, reviewed by people who understand the work, will reveal a lot.
Be specific about what a contract locks in
A longer agreement is not automatically a bad decision.
I would want clear answers to a few questions:
- Can we use another model where there is a legitimate need?
- Can we move our information and work out of the system in a usable form?
- What happens to our data when the agreement ends?
- Can the provider use our information to train its models?
- What happens if the model we selected is retired or significantly changed?
- Can we review pricing and performance during the agreement?
- How much of our internal process will depend on proprietary features that cannot be moved elsewhere?
There will always be some cost to changing providers. AI models are not interchangeable, and any replacement will need to be tested. The aim is simply to avoid making that change harder and more expensive than it needs to be.
Why being model-agnostic mattered when we built Truekeep
(If you don’t know TrueKeep, it keeps your company’s knowledge, messaging, and voice consistent across AI tools)
We cannot tell a client which model will be best for every kind of work three years from now. I do not think anyone can credibly make that promise.
SO, as we started working more on AI enablement, we wanted to find a way where organizations had one place to maintain information and direction so that any work was not tied unnecessarily to that prediction. Current company knowledge, messaging, voice, examples, etc. is valuable regardless of which model is being used.
If one model is the best fit today, it should be possible to use it. If another becomes better for a particular task, or a client’s security or requirements change, it should be possible to consider that model without rebuilding everything the organization has already put in place.
That flexibility still has to be managed responsibly. Any model change needs review and testing. Different models produce different results. Model-agnostic does not mean treating them as identical.
It means an organization should have room to make a different choice when the facts change.
For organizations making these decisions now, my advice is fairly simple: give employees a useful, approved way to work with AI; be clear about the information and activities that require tighter controls; keep organizational knowledge somewhere you control; test models against your own work; and avoid agreements that make a future change unnecessarily painful.
You do not need to identify the permanent winner in AI right now. You need to make a sound decision for the work in front of you and leave yourself enough room to adjust.
