• Bitter Lesson, Rich Sutton, 2019
  It's bitter because the lesson keeps getting re-taught. Attempts to embed our interpretation of discovery and understanding into AI are quickly surpassed in performance by generalized approaches to search and learning with increased computing power. Otherwise summarized: "the only thing that matters in the long run is the leveraging of computation". In the field of finance and accounting, where we have a tendency to hold in high esteem the contribution of human experience, I wonder how this'll play out in application of AI and agentic AI. It's compelling (and comforting) to think our current methods and processes contain permanent value, but maybe it's all rather irrelevant in the not too distant future? In which case, our current problem solving approaches aren’t useful in training models, but our ongoing interactions are critical for inference.

• Roadmap: AI systems of action, Bessemer Venture Partners, 2025
  An important thread in the disruption of legacy ERP is the application of AI to the implementation process, which "means the time and cost to move to a new system are significantly less." This has big (ooh, huge) potential: Current incumbents' moats rely upon high switching costs, remove that and you have a much more open playing field. And not to mention the increased ease of building features in platforms using AI, this means—as I think has been said most succinctly by Franz Farber of Everest Systems—the true value of platforms will be in the guarantee of an outcome (which is really the value the customer actually needs), e.g. a guarantee your financials are accurate and you will be audit ready…

• Future of Work with AI Agents, Stanford University, 2025
  "Domain workers want automation for low-value and repetitive tasks." People naturally want to spend less time doing the boring stuff and it means more time is available for higher-value work, which should result in higher profit margins. Win-win all around then.
  Though I wonder: Sometimes we enjoy not having to turn our brains on, especially the morning after the night before... do we switch AI off at that point?
  The paper also highlights the three most prominent concerns about AI:
  1. lack of trust in accuracy
  2. fear of job replacement
  3. absence of human qualities or capabilities (see above and the Bitter Lesson)
  Then in conclusion "From an incentive perspective, some workers may also withhold honest feedback due to concerns about job security or surveillance." Key takeaway for executives in companies investing in AI usage—bring the team along and focus on reskilling, they are part of the solution not a consequence of it.

• How CFOs are navigating growth, pricing, and forecasting in an AI world, a16z, July 2025
  Annual subscription licenses and revenue recognition often mask the underlying business health—customers may not even be using the platform, but yet the P&L shows revenue recognition on an amortized basis. This happens more often than you'd think. Consumption based pricing (and consequently revenue recognition) would better tie reported financial performance to actual customer value delivered. Finance processes to efficiently implement this have to depend on automation.
  Approaches to ARR then need refined to include annualized estimates of consumption revenue. And then there’s margin: Optimizing margin in operations becomes a fine-grain task of managing costs with usage, requiring clarity on financial impact on a much shorter timeframe than usual month close cycles. Systems need to reliably link activity to predicted financial consequences so optimizations can be made throughout each month.

  I agree with this: "You can't achieve the precision required in consumption forecasting using Microsoft Excel — you have to use advanced analytics, machine learning, and AI to build those predictions."

  But, the unending flexibility of Excel is going to be very difficult to unseat as the principle work medium for finance, FP&A etc. It's a question in the nearer term then of how Excel is augmented (time to highlight Aleph as a key player) by the increased ease of data access and all that brings to the challenge. And then we're into how AI can help make sense of that data—will be fascinating to see how interfaces will evolve to enable natural language direction.

• Blub's Paradox in Finance, Looks Gerat!, March 2025
  Great perspective on why Excel users tend to stick with Excel. Before LLMs and the advent of easy code generation it took a lot more brain power and time commitment to wrangle Python for data processing, Now it's remarkably easy to craft up a script for manipulating and visualizing data—training LLMs benefits from copious amounts of Python code in the ether and so the resulting models can produce pretty useful code straight off the bat. However, even though complex Excel formulae represent many coding constructs (and so Excel power users should be able to pick up more powerful languages) the whole environment of coding is vastly different to that of spreadsheets. Yes, stuff like version control becomes a built-in feature rather than hacked using filenames, but the user needs to build familiarity with projects, libraries, version control and repos, managing virtual environments, not to mention unit testing—which is snother of the massive benefits of coding vs spreadsheets--and so on to actually access the power. I'm a big proponent of moving large portions of what we deal with in financial projections, modeling and analysis to coding—it's just that the complexity of doing that and maintaining reliability and integrity in the output shouldn't be underestimated.

In the Glass…

Glen Scotia 7 Year, 1st Fill Ruby Port Hogshead

Exclusive cask, bottled for Loch Fyne Whiskies

Picked this up on a recent trip back home, from Loch Fyne Whiskies in Inverary, an almost guaranteed stop for me on the drive west. Glen Scotia is one of my absolute favourite distilleries and I never fail to enjoy what they release (code #93 with SMWS by the way, and I have a few of those bottles). The exclusives that LFW get are always excellent and sell out very quickly. This one is no different, the port notes clear and distinct, and the slight smoke that just trickles through on the palate. Terrific stuff.

Last Thought…

I used em dashes long before ChatGPT got into the game but I take no responsibility for correct or incorrect usage, or indeed for the mixed up British vs US spelling above. However, I do strongly recommend checking out ellecordova’s Instagram for some fun with grammar and planets.

Subscribe now

The potential for AI to automate functions like finance is touted by legacy platforms, emerging technology solutions, investors and even some finance leaders. But it's clear to me that the solutions most likely to succeed all feature "Human in the Loop" (HITL).

The amount of marketing collateral and online discourse around AI in finance is bewildering. Yes, there will unquestionably be huge benefits over time but it can be very difficult to separate wheat from the chaff in understanding what these platforms currently use AI to do, where they actually use AI (you are perfectly within your rights to ask this of vendors and require a detailed response) and then what it means for the user.

Focusing on an HITL approach in a product is pretty sensible–done right it serves both as a practical necessity for accuracy and a psychological bridge for user adoption. Has struck me recently, though, as to whether this is singularly a technical requirement for success or just a clever strategy to better ensure adoption by users.

Is HITL playing it safe or a perfectly honest marketing message?

Walk through any finance technology conference or browse vendor websites, and you'll notice a pattern. Nearly every AI solution emphasizes how it "empowers finance professionals" rather than replaces them. The message is carefully crafted: AI handles the mundane while humans focus on "strategic work." It's a compelling narrative that addresses the elephant in every room–will AI take my job?

This positioning is brilliant from a go-to-market perspective. Finance teams are naturally conservative and risk-averse. They're also the ones who, importantly for the vendor, control the purchasing decisions. By promoting HITL, vendors transform from threats into partners. The AI becomes a helpful copilot or assistant rather than a replacement. The fear factor disappears, replaced by curiosity about efficiency gains.

Perhaps there’s even a pricing element that, paradoxically, it’s possible to charge more for a platform that requires human oversight than one without. This plays to the conspiracy believers that it’s a marketing ploy.

Or maybe this messaging is actually honest.

When humans actually matter: connecting the dots

Working in finance, whether in accounting, FP&A or at the leadership level, frequently requires judgment and subjective analysis. So much of what makes up finance is based on assumptions around known unknowns (eek), topics lacking clarity or application of accounting guidelines to situations that don't easily fit into one box or the other.

Take for example an unusual or limited-information transaction in the month close process–should it be expensed, accrued, or flagged as an issue? A reliable answer to this usually is guided by accountant knowledge, years of finance experience, and contextual information perhaps only available at the business level.

What's also interesting here is how platforms will likely evolve to include a greater degree of business activity and data for AI to review. It's compelling to think of AI accessing vendor contracts, email threads about ongoing negotiations, Slack messages about project status, even calendar entries about customer meetings–all to better determine the nature of a transaction and how to handle it in accounting. The AI could theoretically triangulate from multiple data sources to understand context the way a senior accountant does through experience and conversation–and with much greater speed.

But until we're there–and we're definitely not there yet–it's really only feasible for a human to step in and close the loop. The gap between available data and required judgment remains too wide for AI to bridge alone.

The adoption game hinges on understanding accountants (not just accounting)

But, being honest, the HITL emphasis is also shrewd business strategy. I've seen enough technology implementations to know that the human element often is the determining factor between success and failure. Sometimes there are change-resisters who just throw up roadblocks, or at an extreme people who feel that their job is at risk then deliberately sabotage the project. You’d be amazed at what can happen.

HITL flips this dynamic. When the finance manager can review and override AI suggestions, they feel in control. When they see their corrections training the system, they become invested in its success. Each month, as they trust the AI with more routine decisions, they're actively participating in their own role evolution rather than being dragged along and eventually snuffed out.

This psychological element might be even more important than the technical requirements. A solution that's 90% accurate with enthusiastic users will succeed where a 95% accurate solution with resistant users fails. Vendors who understand this truth position HITL not as a limitation but as truly as a differentiating feature.

Finance as judgement not calculation

Yet still it's tempting to view HITL as primarily a marketing strategy–a temporary comfort blanket until AI gets good enough to fully automate finance functions. There's certainly some truth to this. Basic transaction coding and invoice matching will likely need less human oversight over time. The "human in the loop" for these tasks is partly about building trust during the transition.

But this cynical view misses something important. The aspects of finance that require judgment, context, and accountability aren't edge cases–they're core to the function. Financial reporting isn't just about processing transactions; it's about telling the business story accurately and defensibly. FP&A isn't just about financial metrics and drawing lines on a chart; it's about understanding strategy and market dynamics.

The vendors emphasizing HITL might be solving for adoption, but they're also acknowledging a fundamental truth: finance is as much about judgment as calculation. Correspondingly, it can be asserted that the most successful AI implementations will be those that enhance human judgment rather than attempt to replace it.

So where does this leave us…

I can’t see any other (partial) conclusion than: HITL serves dual purposes, technical necessity and adoption catalyst. Smart vendors recognize both aspects and design their solutions accordingly. They're not being dishonest when they emphasize human empowerment–they're being realistic about both current AI capabilities and the nature of finance work.

I think that for finance leaders evaluating AI solutions, understanding this duality is really helpful. Yes, some vendors probably do oversell HITL to quell fears about job displacement. But I think the best solutions genuinely need human expertise to function effectively in the complex, judgment-heavy world of finance. Going beyond that statement right now risks getting drowned in untethered hyperbole about AI taking over the world and replacing everyone except plumbers (which of course are separately at risk from the robots).

The interesting question isn't whether HITL is necessary or just clever marketing (it's really kind of both). After all, in a field where "it depends" is often a critical component in developing an accurate answer, keeping humans in the loop ensures a better outcome for the user and customer.

And AI’s contribution:

I chucked the title and subtitle into ChatGPT 5 and asked it to generate a radical image. Not bad, and the words are all correct.