Systems of record and transaction in an AI agent world

Coming up roses.

Warning: Long post. Some metaphor overload. Thoughts in progress.

The investing world is abuzz with more agents than Glienicker Brücke on a foggy evening in a spy novel set in the 1960s. GenAI seems so last week. I’ve still not figured out whether LLMs have brought us to the edge of AGI or whether they are fatally flawed, or somewhere in between. I’m learning to be okay with being unsure.

In the last couple of weeks I’ve probably received 20 pitches for recruitment AI agents which promise end to end automation and more. It may well be that someone gets this right, but is hard to pick one, but we probably should.

Agent is a rather interesting choice of word. Agentic is a grammatical abomination, but I suspect we are stuck with it.

In law, an agent is someone that acts on your behalf, and agents have specific duties, obligations and so on. As we delegate more work to these agents, it’s going to create some very interesting questions, not just in terms of data protection law, but in fundamental contract and liability. I’ll dodge that bullet for now, but read what Rachel has to say.

My former colleagues at Gartner are predicting that, by 2028, at least 15% of day-to-day work decisions will be taken autonomously through agentic AI, up from 0% in 2024. (via Nico Orie’s post)

The larger VC firms and the usual strategy consultancies helpfully write lots of collateral that can “help” us decipher what they think agents are and will be. I read their work, some of it is really insightful, but I always bear in mind the incentives of those writing it (as you should reading what I write). Lately I’ve found IBM’s content on AI and related developments to be excellent, and the hype levels pleasantly subdued. This was a particularly useful introduction (Interesting that Maya uses an HR system of record use case).

hat-tip Matt Candy from IBM.

Last week or so there was also quite a bit of excitement about what the latest version of Claude can do, or promises to do. Have a look at this job application example.

https://x.com/minchoi/status/1849110828378665356

Nifty demos put together a couple of hours after a product launches are intriguing and eye opening, but they typically don’t actually mean much in terms of genuinely deployable enterprise software.

But talking with Dan at Peoplereign, Jerome at Datascalehr, the folks at ServiceNow and with Jeremiah Stone, the CTO at Snaplogic, agents are now making a meaningful impact in production in large enterprises.

https://x.com/jeremiahstone/status/1845495727763697781

I’m pleased to see us moving to productive use cases, rather than mere pundit proclamations of revolutions. If you get the chance, watch Jeremiah’s talks.

A thought experiment

Firstly let’s assume for a moment that the excitement in agents is completely justified, and that they work really well, filling in (UK) and out (US) forms, booking travel, timesheets, reconciling invoices, finding order discrepancies, discovering prospects, signing contracts etc on our behalf. I’m going to assume that any legal agent liability issues are cleanly managed, and its secure and all the boring enterprisey stuff is good. They don’t go rogue.

We will have agents successfully executing all sorts of work on our behalf, updating the systems and processes that we spent decades designing (often not well, for human users). We trust them to get work done on our behalf: Gartner’s prediction lands.

So for the next few years, we will have AI agents running around filling in forms and fields that were designed for humans and for traditional relational database structures, derived from client server and first generation SaaS. Basically delivering on the earlier promises of RPA. This is going to change a lot of the processes for support and service, for the better, but it is innovation on top of an existing data model and application. It may shift value creation, destroy and create new jobs, but it isn’t yet fundamental disruption.

So here’s the experiment question. If we (with the help of AI) were to completely rethink the underlying transactional and application systems for a world with agents, what would they look like?

Today we think of structured and unstructured information. Structured data is structured so that it can be stored and processed effectively by a computer. Most structured data is an artifact determined by the limitations of what the computer needs to do its job. Early computers stored yes as 1 and no as zero, or male as 1 and female as 2, mainly for efficiency reasons. Numeric fields are cheaper to store and sort, and back then that was a big deal. So we reduce information about people into a series of numbers and values that we can store in data models designed by our parents or grandparents. For instance, the data models we have today are not significantly different from those that supported the first payroll project. The architect of the first payroll would be able to understand today’s payroll rules engines and table structures.

Once we figured out the rules of normalization HR data models haven’t changed much. Rules engines have become easier to use, but they still work much the same way as they did 50 years ago. We awkwardly take unstructured data, and move it into a structured form so that a computer can work with it, using a predetermined set of rules.

When humans write, think and communicate, they typically don’t normalize data, or have drop down table entries. Human processing and record keeping is largely unstructured. We internalize rules and processes, and build clever heuristics to remember things.

In a world where we can trust AI agents, why would we need to manually convert HR regulations and policies from document form into explicit tables and formulae? We could simply ask the agent to collate and read the regulations itself and figure out how to calculate the payroll or the leave policy. In essence, outsourcing it. Why would we turn an employee performance discussion into form when we can distill its essence?

The more human-like AI becomes, the less like a computer database it will need to be. As we and AI learn, we and AI will fundamentally rethink what transaction processing and system of records need to be.

My conjecture coming out of the thought experiment is: the more initiative we give to AI agents, the smaller systems of transaction and record will need to be. They become less relevant, not quite trivial, but almost.

I’m not the only one pondering this, have a look at Nico’s and Martin’s posts.

The corollary of this conjecture is that auditing AI is not going to be trivial.

A background metaphor

I’ve a metaphor about flowers that I have used for about a decade to describe HR systems architecture and ecosystems. I’ll repeat it here (see earlier post for more).

“Large organizations have different cultures and behaviours when engaging with software companies, and I have a metaphor I use to explain this.

The sunflower: large core, small petals. These organizations commit to a suite vendor as their dominant provider. They have to use the suite unless it clearly is unfit for what they need. IT is normally quite powerful, and factors such as integration, suite vendor partnerships, joint GTM, certifications etc are important. Start-ups will need to collaborate with the suite vendor to be successful in these companies.

The daisy: small core, large petals. These organizations use their suite vendor as the system of record, but they rely on niche vendors for innovation. They are easier to sell to, especially at the LOB level. Start-ups compete by differentiation, not integration. Winning deals is easier, but gaining enterprise wide traction and stickiness is harder.

The cactus: This is an anti-pattern. These companies only use the suite, even if the suite fails to meet their needs. These organizations are impossible to sell to. They have lots of shadow processes in Excel and the like.

The dandelion: This is the most dangerous anti-pattern. These companies very quickly pilot software, do lots of POCs, but they never actually deploy stuff. They waste the time of startups and they frustrate their end users with lots of half baked stuff. I see these companies on every pitch deck.

My advice to large organizations is be clear who you are. If you are a sunflower, you will have long sales cycles but you are prepared to commit for a long and meaningful commercial relationship. If you are a daisy, you can move fast to buy, deploy and replace. If you are a cactus or dandelion, get therapy.

Adding a new flower to my bouquet

It might be time to add a rose to the bunch. Unlike the sunflower and the daisy, the rose is almost all petals. Think of agents and new AI first applications as the petals. Part of the beauty of the rose comes from how the petals are layered together. The petals of the rose are seperate, yet precisely coordinated. The sum is greater than the parts. We may see multiple agents, collaborating in a coordinated fashion, shielding us completely from the core (I have not figured out the thorns yet).

So who wins this in this AI centric world?

There is a lot that needs to go right for the system of record and calculation to shrink dramatically, but if it does, the massive structured databases and rules engines of today will be obsolete.

The ecosystems in HR tech have evolved over the the last 25 years or so, since the heyday of client-server, but they haven’t been really disrupted. For all the talk of disruption to date who is really switching off their incumbent products (I don’t really buy the Klarna story btw)? The scope of HR tech has growth dramatically, the basic underlying choice of sunflower v daisy has remained consistent. It’s time to shake this up.

Most management theory would argue that insurgent vendors will disrupt the status quo, and we expect to see new players developing. But the incumbents have an opportunity to continue to thrive. For instance in payroll, incumbents like ADP, UKG, SDworx and Dayforce have a significant opportunity to innovate, in part because they have access to decades of training data. But disrupting decades of existing business process and product will not be easy. ServiceNow is in one sense an incumbent, but it is also an insurgent outside of ITSM, and it is well placed to benefit in an agent centric HR world. The size of its HR business dwarfs most of the last decade’s HR tech darlings.

At Acadian we think there also will be new set of vendors joining the fray.

If you are growing a rose, I really want to talk. In a future post I’ll dig more into the details of what I think the HR stack looks like over the next decade. I’ll also apply some more ecosystem theory from Michael Jacobides.

In the meantime, let’s take you back to 1987. It’s a goody.

And a lovely bonus song from 2001. This one makes me smile, whimsically.