AI is being bought for speed, scale and efficiency. Every company that buys it will get those, eventually. The more interesting effect is one nobody is paying for.
Widely adopted technologies have trended towards making their adopters more similar. A capability only some firms can obtain is an advantage; a capability every firm can obtain may be enormously valuable but distinguish nobody.
Nicholas Carr made the point sharply in 2003, in an essay that annoyed most of the technology industry and has aged better than its critics: as information technology becomes ubiquitous, affordable and standardised, it stops conferring advantage and becomes a routine cost of doing business. If he was right, and two decades of cloud and enterprise SaaS suggest he was, the result is that process innovation suffers as a means of gaining competitive advantage.
AI is currently being adopted for benefits of exactly the equalising kind: throughput, cost, scale, speed, and controversially less human effort per unit of output. The predictions are significant, and they are why almost every deployment is funded. They are also, on Carr's own logic, temporary. Everyone will get them. Which means the durable AI competitive advantage has to come from somewhere else.
Except that this one is categorically unlike the technologies that produced Carr's pattern. This technology arrives with the ability to also produce more technology — workflows, applications, processes, agents that run unattended. The pattern of SaaS was to generalize best practice and resell it, so its adopters converged on one artifact. A tool that can both execute and create lets each adopter make something different.
So the benefits split in two. The efficiency gains are shared and parts of the advantage will commoditise on the usual schedule. What gets created or enhanced by the technology is not shared, because it is made locally, with different context and different ideas.
This is where a business can find a significant advantage, and it changes what adoption costs. Previous technology adoption made a company more efficient and less distinct at once, and the loss of distinction was accepted as the price of the efficiency. AI adoption changes that.
Why did past technology trade differentiation for efficiency?
Cloud and enterprise SaaS delivered capability no individual firm could have funded, and a standard that was legible to auditors, to new hires and to an acquirer, and frequently better than what it replaced. The value the platform companies captured was large, and it was payment for value delivered.
Jos Benders, Ronald Batenburg and Heico van der Blonk traced a Dutch publishing company that deployed enterprise resource planning with the explicit intention of differentiating itself, and watched it reshape the organisation to fit the software instead. They named the mechanism technical isomorphism. The standard made one process efficient and every other process expensive, reducing differentiation in favour of compliance.
Adopting the standard was safe and defensible within and outside the organization, further diminishing incentive. Your skill with a tool became a transferable individual advantage in the market, which reflects the ubiquity everyone collapsed to.
This bargain was to the vendors' advantage, entrenching software choice as the primary decision in front of an organization.
SaaS adoption was about compliance, AI adoption is about describing
Instructing an agent means describing the work in your own words, with your own documents, your own exceptions, your own account of what matters and why. None of that is a menu, which is why two people building what is recognisably the same thing end up with different results.
Configuring software meant choosing among options somebody else enumerated — an option space finite, published and identical for every customer, which is why two companies configuring the same CRM end up in the same place. The settings page was the outer limit of how different you were permitted to be. That is a compliance exercise, and most companies were right to accept it.
A form collects your answers. A canvas takes your judgment.
For years the executional layer of a business was a form, refined across thousands of customers, and bounded by what its designers generalized. Agent development behaves more like a canvas. It rewards whatever gets brought to it. The ceiling used to sit inside the product. It now sits with whoever is describing the work.
How does process innovation create AI competitive advantage?
Some business advantage over competitors comes from what they sell. Another advantage comes from how they run. The distributor whose fulfilment is a day faster. The insurer whose underwriting catches what others miss. The retailer whose logistics optimizes at every end.
Those advantages live in process, and under the platform paradigm they were expensive to keep. That is the trade that has changed. The efficiency no longer arrives attached to a requirement that you do things the ordinary way.
That the ideas exist is not a hopeful assumption. Alan Felstead and colleagues, writing in the British Journal of Industrial Relations, set out to test the traditional view that innovation belongs to a small band of highly skilled workers, and found employees across occupations and sectors routinely coming up with ideas for improving the processes they use and the services they provide. What predicted whether those ideas surfaced was not talent but the conditions around them: involvement, support, development. Their uncomfortable finding was that several of those conditions had declined in Britain over the same period that productivity growth stalled.
The innovative improvements available to an organization, rife in the minds of the workers manning those workflows, can be readily executed as part of AI adoption. An agent platform that delivers reliable workflow creation through the same conversational paradigm as AI chat opens process innovation to the people who have always understood the process best, and who have never had an affordable route from an idea to a working change.
Do AI workflows replace standards or build on top of them?
Agents can create workflows of their own or reach into the traditional systems, connecting them, enhancing them, and making the result specific to you. Differentiation moves up a level, from which systems you run, which is roughly the same list everybody runs, to what you do across them.
In a world of AI, the advantages of standards — shared visibility, shared definitions, shared process — are no longer something necessarily purchased. A standard becomes a way agents interact, something that stacks on top of new forms and workflows rather than something those workflows are bound by.
What separates AI agents from ordinary automation is what sits underneath them. It keeps a record of its own runs, so there is something to look at when it behaves oddly. It carries context: your documents, your exceptions, your account of what matters. It can be scored, by a person marking the output or by a check that runs on its own. And it can be changed by the person who wanted it, in the afternoon, without anybody's permission.
So the work is not fixed at the moment it is set up. It accumulates your specifics as it runs, because running it is how you find out what you actually meant. It is also how you get out from under the model's average: context is the thing that pulls it, and nobody else has yours.
Why do small workflow changes add up over time?
Every organization knows more than it can use. Some of that knowledge is written down — policies, playbooks, past decisions, the notes attached to an account, the document somebody wrote in 2021 that turned out to be right. Most of it is not written down at all. It sits with the people doing the work, in the form of what they have learned by doing it.
Neither form was ever executable. A document describes how something should go; a person does it. Nothing in between could carry both.
An agent takes both. You give it what is written down, and you tell it what is not — the exceptions, the cases that need a second look, the reason a step exists that nobody ever documented. That is the starting context, and it is already more than any configured system has ever held about your business.
Then it keeps developing, which is the part with no precedent. Each run adds to the record: what came up, what was decided, what failed and what was done about it. That is organizational knowledge the company never captured before, because it evaporated at the end of the day or left the building with whoever held it.
Memory is what makes that durable: the exception that surfaced in March is still there in September, along with what was done about it. Reusable pieces mean an improvement made once can be called from everywhere — a rule about which suppliers need a second look, written for one workflow, becomes available to every workflow that should have it. And evaluation means an improvement can be checked rather than argued for: a change either scores better than what it replaced, or it does not.
Together these close a loop. The process runs, the run is recorded, the result is scored, somebody adjusts, and the next run starts from a better place.
What is the cold start problem in AI adoption?
SaaS and platforms made starting out of the box an enviable benefit. Investment in UX and the advantage of standardized familiarity made initial impact measurable. But as usage grew over time and cases, configuration, features, specificities, definitions and data organization escalated. The burden of its limits arrived later, when organizations tried to customize, scale, or innovate.
AI and agents demand more from its users up front. Chatting with AI for an ad hoc output has become rapidly intuitive, even then it still takes iteration. Describing a workflow is a larger commitment, with more iterations and more tangential steps — connectors, keys, documents. The industry is responding by re-integrating elements of UI and offering templates that stay malleable once opened. Larger organizations are staffing against that first climb. Smaller ones are getting the same effect from a platform that asks for a description rather than a specification.
The effort is measurable. BCG's 2026 workplace survey, covering 11,749 workers across fourteen markets, found nearly half of employees now spending more time managing and directing AI than doing the work themselves, and 41% reporting higher cognitive load alongside the gains.
The shape of the cost has moved rather than disappeared. It used to arrive late, when you wanted to be different. Now it arrives early, while you are still working out what to initiate.
What does building an AI competitive advantage ask of a company?
Advantage accrues to whoever can articulate their own work, which is a skill and unevenly held. More attempts is not the same as more good attempts. And a process shaped this specifically is harder to audit, to staff and to hand over — which is the property that makes it hard to copy, seen from the inside.
But the direction is the encouraging part. For years the way a company did things was largely a function of what it had bought, and the ambition of the people running those processes had nowhere obvious to go. Competing on process meant paying to stay different.
It no longer does. As existing workflows settle and the process becomes comfortable, the innovation waiting in every corner of a business has somewhere to go. The optimizations that were previously ignored, the initiatives that never got started, the novel approaches that failed to find a landing strip: all of them are now supportable, testable, and tunable by the people who thought of them.
Why does supported innovation outlast any single improvement?
In 1993 Paul Geroski, Stephen Machin and John Van Reenen examined UK firms expecting to find that innovations made money, and found something subtler. The profit from any single innovation faded. Yet firms that innovated stayed persistently more profitable than firms that did not, in a way the innovations themselves could not explain. What paid, on their reading, was something about the innovating firm rather than the innovation.
There is also a reason to expect people to want this. Amy Wrzesniewski and Jane Dutton named it job crafting in 2001: employees changing the characteristics of their own work rather than waiting for a redesign to be handed down. Cort Rudolph and colleagues, pooling 122 samples covering some 35,000 workers, found notable relationships between job crafting and engagement, job satisfaction and performance — though the strength and direction vary by which kind of crafting is involved. That paper describes job crafting explicitly as an individualized, bottom-up approach to job re-design, set against the top-down, one-size-fits-all approaches initiated by the organization — which is a fair description of a platform standard.
The current evidence points the same way. In that same BCG survey, two-thirds of regular AI users reported improved job satisfaction. People craft their work where they have the autonomy and the means, and for the first time both are within reach of the person doing the work.
No single change is impressive on its own. The accumulation is, and it is difficult to copy, because it was assembled out of knowledge derived, honed and held inside one organization.
The ground also stays open longer than it looks. Clayton Christensen's argument about incumbents was that they fail through good management rather than bad, by allocating resources to the articulated demands of their best customers. A platform business is bound by something stronger than attention. Its economics depend on customers being alike: the support model, the upgrade path, the documentation, the sales motion. A vendor can add configuration, and will. What it cannot rationally do is make every customer's process different from every other's, because that is the same as dismantling the standard it sells.
That constraint applies to a vendor selling a finished process. A vendor selling the means to build one has the opposite incentive, since its customers diverging is what proves the thing works.
Michael Polanyi wrote in 1966 that we can know more than we can tell. The line is usually read as a limit on people. It is at least as much a limit on what was listening. Software wanted rules — stated in advance, complete, with the exceptions enumerated. People hold their work as examples, conditions and usually-except-when-this-happens. Nobody was refusing to explain. There was nothing that could take the explanation in the form it arrived in.
That is the change, and it is why the cold start is worth paying. Describing your work is real effort. But it is effort of a kind people can actually make — talking about what they do — rather than translation into a format they were never going to master.
If how you do things is what makes your business better, this enhances that rather than diminishing it. The scaling benefit of AI remains. Come for the efficiency. Stay for the innovator advantage.
Postscript: visible innovation accruing on an agent platform.
At CREAO we have the privilege of observing the growth of innovative thinking occurring in real time. This comes from our own platform, where a wide range of early adopters are creating and growing their own AI agents.
Among early adopters we see a landscape of innovations: some finding new ways where an app may already exist, some connecting applications in unique value-add ways, and others supporting workflows nobody else runs.
Of the agents running, three-quarters carry a name no other user has ever used. These are not drafts or abandoned experiments. They are workflows somebody relies on, and they are very often one of a kind.
They also keep changing. A third of running agents carry a version number above the one they started with, and the changes arrive after weeks of running rather than during setup. In the old paradigm that refinement was expensive. Now process improves at the pace of outcome.
Scaling with AI will become expected. Innovation within that will grow as an advantage.
→ Start building your process advantage
References
Benders, J., Batenburg, R. and van der Blonk, H. (2006). "Sticking to standards: technical and other isomorphic pressures in deploying ERP-systems." Information & Management, 43(2), 194–203. doi:10.1016/j.im.2005.06.002
Boston Consulting Group (2026). AI at Work: Why Strategy Matters More Than Tools. Fourth annual global survey of 11,749 employees across 14 markets, June 2026.
Carr, N. G. (2003). "IT Doesn't Matter." Harvard Business Review, 81(5), May 2003, 41–49.
Christensen, C. M. (1997). The Innovator's Dilemma: When New Technologies Cause Great Firms to Fail. Harvard Business School Press.
Felstead, A., Gallie, D., Green, F. and Henseke, G. (2020). "Getting the Measure of Employee-Driven Innovation and Its Workplace Correlates." British Journal of Industrial Relations, 58(4), 904–935.
Geroski, P., Machin, S. and Van Reenen, J. (1993). "The Profitability of Innovating Firms." RAND Journal of Economics, 24(2), 198–211. doi:10.2307/2555757
Polanyi, M. (1966). The Tacit Dimension. Doubleday & Company. (Reissued by University of Chicago Press, 2009.)
Rudolph, C. W., Katz, I. M., Lavigne, K. N. and Zacher, H. (2017). "Job crafting: A meta-analysis of relationships with individual differences, job characteristics, and work outcomes." Journal of Vocational Behavior, 102, 112–138. doi:10.1016/j.jvb.2017.05.008
Wrzesniewski, A. and Dutton, J. E. (2001). "Crafting a Job: Revisioning Employees as Active Crafters of Their Work." Academy of Management Review, 26(2), 179–201.
Related reading
FAQ
Q: What is the difference between AI efficiency and AI innovation?
A: Efficiency is the throughput, cost and speed gain that arrives with the technology. Every company that adopts AI gets some version of it. Innovation is what gets built on top: workflows, agents and processes shaped around one organization's context. The first is shared. The second is made locally and stays local.
Q: How do AI agents create competitive advantage?
A: An agent carries context that no configured system held before: your documents, your exceptions, your account of why a step exists. It keeps a record of its runs, it can be scored, and it can be changed by the person who wanted it changed. Two businesses building what looks like the same agent end up with different results, because the inputs were never the same.
Q: Why does AI adoption have a cold-start problem?
A: Describing a workflow takes more effort up front than configuring software did. There are iterations, connectors, keys and documents before anything runs reliably. Older platforms front-loaded the ease and charged later, when you wanted to do something the standard did not allow. Agent platforms reverse the order. The cost arrives early, while you are still working out what you meant.
Q: Does AI adoption make companies more similar or more different?
A: Prior waves of enterprise technology made adopters converge, because each one generalized best practice and resold it. Agent platforms behave differently, since the same tool that executes work can also create new work. Differentiation moves up a level, from which systems you run to what you do across them.
Q: Who inside a business is best placed to build AI workflows?
A: The people already running the process. Research on employee-driven innovation finds ideas for process improvement spread across occupations rather than concentrated in specialist roles. Conversational agent building removes the translation step that used to sit between an idea and a working change.