The AI Differentiation Dividend
Almost every team now uses AI. Very few enterprises can show a predictable return on it.
Heading into the second half of 2026, that is the pattern I keep seeing, in travel and well beyond it: token spend climbs, but the increment is uneven and rarely company-wide. A few teams pull real gains. Most do not. An MIT study put it bluntly this year, finding that 95% of enterprise generative-AI pilots delivered no measurable return.
My view is simple. The advantage will not go to whoever deploys the most AI. It will go to whoever has the system to turn it into predictable output they can trust, and then spends their best people on what actually sets the business apart. That, to me, is the differentiation dividend.
More spend, uneven return
The honest truth is that more AI has not made most companies a penny. Adoption is near universal and spend keeps rising, yet the return stays patchy. 2026 telemetry from Faros shows the strain: as AI lifts output, a third more changes now merge with no review at all.
Here is why. We reach for AI the way we already use chat tools, as a quick fix bolted onto today’s work. That builds a faster horse, not a new way to move. Pour AI into a weak process and the speed leaks straight back out as fragility.
The variable was never the model. It is the system around it. What turns scattered, team-by-team wins into a reliable, company-wide increment is not more AI. It is the system.
What actually decides whether AI pays back
DORA’s 2026 research on AI ROI, the work I lean on most here, is blunt about it. The greatest returns come not from the tools themselves but from the underlying organisational system: the quality of your internal platform, the clarity of your workflows, and the alignment of your teams. AI is an amplifier. It magnifies the strengths of a high-performing organisation and the dysfunctions of a struggling one. Without that foundation, AI creates localised pockets of productivity that are lost in downstream chaos.
Five parts of that system decide whether it pays back. Point AI at a real problem, because user-centricity is the top prerequisite for any return. Give it a high-quality internal platform to build on. Put safety nets around speed, because AI lifts throughput but strains stability, and testing and fast feedback are what hold the line. Keep teams healthy, because well-being and psychological safety still predict who pulls ahead. And measure what matters, people and flow, not lines of code.
Performance, in other words, is a property of the system, not the tool. Which is exactly why AI amplifies rather than fixes.
Automate the ordinary, invest in the extraordinary
The way I think about it, every hour of work sits on one side of a line. On one side is undifferentiated grind that no traveller ever thanks you for: profile and data syncing, rate and fare auditing, confirmation chasing, content normalising. On the other is the differentiating judgment and relationships that earn loyalty: revenue optimisation with the right offer at the right moment, true personalisation, service, and the loyalty that earns the next trip.
I would hand the ordinary to agents, but keep people in the loop on both sides, judging at the right moments. The scaffolding is commodity now; you can vibe-code it. The point is to spend your scarce human attention on what truly sets you apart.
The upside is real, but it is earned. The teams reaching 3 to 10 times acceleration on structured tasks are the ones that rewired the work, not just added the tool.
Trust is the quality of the output
Trust is subtle. It is the quality of the output, nothing more. And that quality rests on context the system has to gather twice: enough to write the right code when you build AI, and enough to read what a user actually meant when you run it. The same discipline on both sides, and both are easy to get wrong.
Context is fragile. As work scales, that understanding degrades, and the output starts to look right while being wrong. It slips in four quiet ways. Context compresses: without sub-agents to divide the work, an agent crams everything into one window and, as it fills, quietly gets less capable. Meaning gets misread: without the right context, “Naples” could be Italy or Florida, and one wrong read sends a confident answer to the wrong place. Retrieval misses: pull the wrong document and the agent answers fluently from the wrong facts. And it cannot judge its own quality: trained on a code repository of good code and bad, an agent copies what is common, not what is right. Telling craft from technical debt is the human’s job.
For the business, the consequence is plain. Confident, wrong output is what erodes customer trust and inflates cost. Humans in the loop are the safeguard.
You can measure AI trust
You cannot measure human trust, but you can measure AI trust, and that changes everything. We do not treat AI quality as a launch checkpoint. For us it is a standing question. We build observability and governance in from the start and keep measuring, asking three questions we never stop asking. Is it right today, not just running? When it is wrong, where and why? Did the last change make it better, or just look better?
This is not theory. On a live travel AI assistant, continuous, attributed measurement took grounded accuracy from the low-50s into the high-80s, and cut cost per answer by roughly a third. Proven quality then let us route routine work to cheaper, smaller models. Reliability and efficiency moved together.
Measurement is not a tax on shipping. It is what turns “is this safe to deploy?” into “how fast can we scale?”
What it is worth to the business
Better ways of working and measured trust do not just speed up engineering. They produce predictable, trustworthy output that shows up in the metrics the business runs on.
You ship safely and faster, with more tested and shipped with confidence and quicker revenue and RevPAR experiments. You get grounded answers, which means fewer wrong outputs, lower service cost, higher CSAT and NPS, and fewer refunds. You run right-sized models, putting routine work on cheaper, smaller models you have proven safe, so token and inference cost fall. And the whole thing is observable and governable, measured, attributed and auditable, which is exactly what compliance and duty of care demand of enterprise buyers.
The payoff is not faster AI. It is a business that can trust what it ships, and prove it.
The dividend
Get the system right, and delivery becomes predictable. Then you are free to focus on what makes you, you.
That system is what my team and I build: solution design for an Agentic SDLC, an AI Centre of Excellence, and Artisyn. The approach is deliberately unglamorous. Start narrow, on a bounded, high-value workflow and outcomes you can measure. Wrap it in observability and governance before it touches a customer or a policy. Prove the business metric, then earn the right to scale across the estate.
The businesses that win the AI era will not be the ones with the most AI. They will be the ones with the system that turns it into a predictable, trustworthy increment, and the discipline to spend the dividend on what only they can do.
If that is a problem you are looking to solve, I would like to hear how you are seeing it. Let’s talk.
Sources:
- DORA (2026), ROI of AI-Assisted Software Development — dora.dev/ai/roi/report (also via Google Cloud)
- MIT NANDA (2025), The GenAI Divide: State of AI in Business — report PDF (coverage: Fortune via Yahoo)
- Faros, AI Engineering Report 2026 (The Acceleration Whiplash) — faros.ai/research · report PDF · takeaways
- SPACE framework (Forsgren et al.) — queue.acm.org/detail.cfm?id=3454124
- Team Topologies — teamtopologies.com
- DataArt — dataart.com
Download the PDF
