There’s so much AI work to be done

Aharon Logue · October 25, 2026

Many early AI prognosticators did a lot of headline-grabbing prognosticating about AI’s impact on work (Goldman Sachs) (knowledge workers decimated! Jobs dropping like flies!). So far, the worst of the predictions have not quite played out. (Yale Budget Lab; Brynjolfsson et al., “Canaries in the Coal Mine”) AI is still diffusing into work broadly and fast, so far as technological diffusions go (Bick et al., “State of GenAI Adoption”; Harvard–St. Louis Fed, “GenAI Adoption Tracker”), but the idea that getting AI into our work will be quick, simple, or inevitable, is not, AI economic researchers seem to think, quite how all that works. (Brynjolfsson et al., “Productivity J-Curve”; UK DSIT, AI Adoption Research)

Getting AI into work is work. And, potentially, quite a lot of work. To access the benefits of AI, time, effort, and employees must either be hired or redirected to AI implementation and maintenance projects. (Brynjolfsson et al., “Productivity J-Curve”) Which makes you wonder: how much work does it take?

“Work” as a concept is a bit bigger than I can handle in a quick blog post, so instead I begin this series on the work of getting AI into work by focusing on a territory I have some experience with, corporate finance. I’m familiar with corporate finance transformation efforts from my time as a researcher in Gartner’s finance practice, and in the below effort estimator tool, I include a representative set of key finance AI projects, along with grounded (if conservative) estimates of the work involved and time saved.

please follow this link for a less cramped viewing experience

The estimates themselves are based mainly on recent case studies for various finance AI projects, coupled with baseline per process time investment benchmarks from polls asking finance teams how much time is spent on various activities. (APQC, Finance Staffing Benchmarks) My selected projects lean well-established so far as AI initiatives go, though there is no doubt the finance developers at Meta and Tata (Steel [Tata Steel, FY2024–25 Integrated Report, 208–09], Play, etc, a lot of the Tata companies had solid finance teams) and Mellon (Microsoft, BNY case study) are cooking up some immensely clever use-cases I am not including here. This is, as with the conservative estimate numbers, a ploy to go out on as few limbs as necessary, even so far as generalized effort estimators are concerned.

One other curious caveat worth noting as you look around the effort estimator: nearly all of the selected projects were automatable in some form before AI. (UiPath, “Finance and Accounting Transformation”) AI may change the calculus on prior automation opportunities, allowing ‘laggard’ teams to catch up to a new ‘laggard’ baseline. But, for laggards and leaders both, AI isn’t free. Someone has to do the AI work to make AI workable, and capacity-strapped teams unwilling to make additional investments may find themselves not dramatically better off than they were when procrastinating on the last set of technological innovations (cloud ERP, RPA, etc).

Another possible surprise is how much lower the time-savings estimates are in this estimator compared to some of the claims made generally, and even some of the benefits occasionally demonstrated. E.g., it is fully possible to achieve 90%+ time savings from an Accounts Payable automation, assuming reasonable transaction volume and a poorly automated starting point. (Microsoft, Pamarco case study; Deloitte, Zora case study) But, a theoretical upper limit may hide the medium-term costs of, for example, incorporating AI review and correction processes into team workflows, or fixing mistakes and adjusting policies. The estimates here are lower in consideration of the totality of work involved, but individual organizations do achieve outlier efficiency gains above the listed ranges.

One final element worth noting: someone has to do work for AI to get into work, but it’s not yet clear who that someone is. The roles of established vendors, cloud ERPs, new SaaS start-ups, and consultants more generally have not fully established themselves in this rapidly changing technology ecosystem. Although many teams will pursue AI projects internally, many will also delegate much of the work of getting AI into work to the all-too-eager forward deployed engineers and AI experts being hired across industries. (OpenAI, “What FDE Means”; “Frontier Alliances”) Although the best strategy is still unclear, at Gartner we often found teams with internal IT capacity were far more capable of accessing and leveraging available process improvement opportunities.

Ultimately, what I believe this estimator conveys is that AI projects are an investment. Like with the steam engine, the biggest benefits came from reorganizing the factory floor, changing the whole of the work around the new opportunities the technology presents, rather than bolting them on to a particular step in the process. (David, “Computer and Dynamo,” 24–26) But, at the end of the day, someone had to move the machines, and decide where they go, and build the process and system to monitor and maintain the new processes.

AI is very new, and it’s not yet clear what reorganizing the factory floor really means. Many of these projects may be, in the scheme of things, less reorganizations of the factory than retrofits of old machines to a new power source (not to butcher the metaphor). For the time being, however, we are all left wondering how AI can be helpful as it is, given where we are, and what it takes to make those benefits possible.

Explore the finance estimator

References