Skip to content

Back to blog

Where does your company sit on the AI adoption frontier?

Most companies aren’t short of AI ambition but the cracks start to show with AI in production. Here’s a simple way to see how much your business is leaving on the table, and what’s keeping it there.

AI is becoming increasingly capable and companies far, wide, and centre are rightfully looking for ways to apply it to their own businesses. However, not every role can be replaced and how different roles can be augmented depends heavily on their human element. Creativity, judgement, and action still require humans either in or on the loop of agents to make sure the right outcome is achieved. So, every business and / or role has a ceiling on what AI can do for it but what does that mean for you?

The AI adoption frontier

We’ve plotted that curve and called it the AI adoption frontier.

Illustrative. Elvar analysis.[1]

At one end of the spectrum, there are business administration tasks that are highly automatable by AI. Invoice processing, reconciliations, data entry, scheduling, and first-line customer queries are structured, rules-based and repeated thousands of times a month. The inputs are documents and systems AI can already read, and the output is easy to verify with deterministic checks. With little human-only work left over, the ceiling on AI’s gains is high.

On the other end, you have the creative and performing arts. As much as it may be fun to watch AI-powered robots dancing on video, they cannot (and probably never will) capture the human element of creativity and emotion going into performances. AI can help around the edges, scheduling rehearsals or editing footage, but what audiences pay for is the performance itself. Hands-on trades sit nearby. An electrician can use AI to plan and quote, but the job is still done by hand.

Most other roles sit across the middle, where judgement, relationships and being physically present are key parts of their functions. A lawyer hands research, first drafts and document review to AI but keeps the advising and the responsibility for it. The same goes for accountants, consultants and account managers. The biggest wins come from automating the routine half of the job so people spend more time on the half only they can do.

Where most companies actually sit

In practice, few companies are anywhere near their own personal frontier. A March 2026 survey of 650 enterprise technology leaders found 78% had at least one AI agent pilot running, but only 14% had reached production at scale.[2] AI is being tried everywhere and trusted almost nowhere. You’ve probably seen it firsthand. Stuck pilots for various tools, little reaching production, and the gains halted at a fraction of their potential. Every new model or tool has to clear risk, contracting and technical onboarding, one vendor at a time.

Illustrative. Elvar analysis.[1]

The distance to the frontier shows up in two familiar patterns. In large organisations, adoption tends to be shallow. A chat assistant for everyone, maybe a few pilot tools for specific ‘priority’ teams, and the everyday processes that you wish were automated get left untouched. In smaller, faster teams, adoption is real but ungoverned. People get the freedom to connect agents to company systems themselves, gain quickly, and create risks to the company that they have no visibility over.

Both leave value on the table. The first never reaches the automatable work. The second reaches it in ways the business can’t stand behind, so the gains stay local and fragile.

What keeps companies from the frontier

We’ve already established that the underlying models powering AI and AI agents are good enough.[3] So why is there still a gap? The answer lies in scaling pilots into full production solutions that teams can then use safely. A June 2026 Levelpath survey of enterprise buyers found 72% of AI purchases over $10,000 took 11 weeks or more, against 57% for traditional software.[4] Security review was the checkpoint most likely to cause delay, coming from three directions at once:

  • Risk: Where does our data go, and who can see it? Is it used to train the vendor’s models? Which sub-processors touch it, and in which countries? Risk teams are asked to sign off on systems whose behaviour isn’t fully predictable, using frameworks built for software that does exactly what it is told.
  • Contracting: Every vendor brings its own terms. Legal has to negotiate data processing agreements, liability caps, ownership of outputs, indemnities and audit rights, often against terms the vendor can change. Multiply that by every model a team wants, and the legal queue becomes the AI roadmap.
  • Technical: Then someone has to plug it in: single sign-on, API keys, which systems the tool can reach and with what permissions, logging what it does and being able to switch it off. Each vendor does this differently, so each integration starts from scratch.

These are the right questions, but they are asked one vendor at a time, by hand, and answered in email threads and spreadsheets. With one model, that is slow. With the portfolio of models that makes AI worth having, it is unworkable, so companies stop at one vendor or stop at the pilot.

What to do next

Ask yourself two questions. How much of your work is genuinely automatable? And how much of it is AI actually doing today? The gap between the two answers is your distance from the frontier.

If you run operations, technology or risk, plot your own business. Take your five highest-volume processes, estimate how much of each is human-only, and compare that with how much AI handles today. Where the gap is widest, ask which of the three blockers is holding it back.

It’s for reasons like these why we built Mantle as the real-time governance layer between your organisation and the AI models and agents it uses. Instead of running risk, contracting and technical reviews from scratch for every vendor, you define your policies once. Which models are approved, what data they may see, what each agent can do, and where a human is needed for sign off. Every new model or agent inherits those rules so onboarding becomes quick, and you can focus on reaching the frontier.

To see where your business sits and what’s holding it back, feel free to reach out to me directly or book a conversation at elvarsystems.co.uk.

Next in this series: the shadow agent, and why it is a different risk from shadow IT.

Work with us.

We are working with a small number of pilot companies on real work, and we are always glad to hear from teams putting agents to use. Tell us what you are trying to move.