In April 2022 we published a piece about where Sofy had come from and where we thought testing was heading. It made three specific predictions about what we would build next.
All three shipped.
| Predicted, April 2022 | What Happened |
| Automatic test case generation using machine learning | Shipped as test authoring agents, tests generated from plain-language descriptions |
| Web testing, extending beyond mobile | Shipped as web testing agents across Chrome, Safari, Firefox and Edge |
| New frontiers beyond the original platform | Shipped as enterprise ERP coverage, SAP and Dynamics 365 module agents |
That is not the interesting part, though it is a reasonable reason to keep reading. The interesting part is that the industry changed underneath those predictions in a way none of us fully anticipated, and the change was not really about test creation at all.
We thought we were solving how tests get written. The actual shift was in who decides what a test should do.
This piece covers what genuinely changed between 2022 and 2026, what got harder rather than easier along the way, and four predictions for 2028 that we are willing to be wrong about in public.
The Shift Nobody Framed Correctly at the Time
In 2022 the conversation was about no-code. The premise was that writing test automation required engineers, engineers were scarce, and the answer was to remove the code so more people could participate. Recorders, visual builders and keyword frameworks were all attempts at that.
It was a reasonable premise and it was solving the wrong constraint.
Removing code lowered the barrier to creating a test. It did nothing about the cost of keeping that test alive, and maintenance, not authoring, is where the money goes. Industry benchmarks put test maintenance at 30–50% of build cost annually, and a recorded test breaks on UI (record-and-playback) change exactly as reliably as a scripted one does. The no-code tools of 2022 made it faster to create an asset that depreciated at the same rate.
No-code made tests cheaper to write and no cheaper to own. That gap is the whole story of the last four years.
What changed between then and now is that the artefact stopped being the point. A recorded flow, a visual diagram and a script are all descriptions of how to perform a test, frozen at the moment they were created. Agent-based systems describe what should be true and derive the how at execution time, which means the description does not go stale when the interface moves.
Mobile-First Became Everything-First
Sofy started in mobile because mobile was where the pain was most acute, device fragmentation, real hardware requirements, an OS release cadence nobody controlled.
What we did not anticipate was how quickly that stopped being a mobile problem. By 2026 the average enterprise release touches a mobile app, a web front end, an API layer and a backend system of record, and users experience the four as one thing. Testing them with four tools, on four schedules, producing four reports nobody reconciles became the normal state, and the boundaries between them became where defects concentrated.
That is why the portfolio expanded the way it did, and the honest framing is that the market moved and we followed rather than the other way round. The interesting technical consequence is that agents covering different layers had to be able to share context multi-agent orchestration, or they were just four tools again with better marketing.
The ERP Turn: When Agents Moved Inside the System
The genuinely unexpected development, and the one that reframed what testing is for, arrived in 2025 and 2026.
SAP embedded Joule agents across S/4HANA, SuccessFactors, Concur and Ariba. Microsoft embedded Copilot agents across Dynamics 365 Finance, Supply Chain, Sales, Customer Service and Business Central. These are not assistants suggesting actions to a human. They interpret vendor emails, code invoices, route approvals and post to the ledger.
Which changes the question testing answers. For thirty years the question was: does the software do what it was built to do? When the software is an agent making judgment calls, the question becomes: is it deciding correctly, is it staying inside its authority, and can you prove either to an auditor?
Testing used to verify that software behaved as specified. Now a growing share of it verifies that software decided well, which is a different discipline wearing the same name.
Most organizations did not choose this. The ERP test agents arrived in a release wave, which inverts the normal governance sequence where controls are designed before deployment. That is why the compliance question surfaced almost immediately, the EU AI Act became fully enforceable in August 2026, and the SEC formed a dedicated SOX enforcement group in March, and neither was written with agentic ERP in mind but both apply to it.
What Got Harder
It would be convenient to present the last four years as problems getting solved. Three of them got meaningfully worse, and two of those are consequences of the thing that improved.
- Knowing whether a test is right. When a person wrote the test, its intent was legible, you could read it. When an agent generates and adapts coverage, verifying that the coverage is correct and complete is a genuinely harder problem, and the industry does not have a settled answer.
- Test data. Traditional provisioning assumes you know in advance what data your tests need, because a human decided what to test. An agent discovers requirements at runtime and will regularly need conditions nobody provisioned test data management for agents. That inverts the model most tooling is built around, and it is the most underestimated line item in any automation budget.
- Observability. An agent can run with zero errors and perfect uptime while making systematically wrong decisions. Metrics, logs and traces were built on the assumption that correct behavior is the absence of failure, and agents break that assumption cleanly. Very few teams have instrumented for it agent observability.
None of these existed as problems in 2022, and all three are direct consequences of automation getting better. That is the usual pattern, solving a constraint reveals the next one, and it is worth saying out loud because a great deal of current vendor messaging implies the constraint stack is empty.
Four Predictions for 2028
The 2022 post made three predictions and put a date on them. Doing that again is the only honest way to write this kind of piece, so here are four things we expect by 2028 that we are willing to be measured against.
1. Agent validation becomes a named role
Right now, responsibility for verifying that AI agents in business systems behave correctly sits between QA, internal audit and the ERP center of excellence, which in practice means nobody. We expect that to resolve into an explicit function with a title, in the same way that site reliability engineering emerged from an ownership gap between development and operations.
2. Vendor-supplied agent governance will prove insufficient on its own
Microsoft already ships simulation and shadow mode for its own Case Management Agent, and Forrester expects half of ERP vendors to launch autonomous governance modules. These are genuinely useful. But a vendor monitoring its own agents is a different assurance proposition from independent validation, and audit practice tends to recognize that distinction. We expect independent verification to become an explicit expectation rather than a nice-to-have.
3. The testing tool category consolidates into a quality platform category
Functional, performance, security and accessibility testing are currently four purchases, four schedules and four sets of results. When one execution can produce all four signals from the same session on the same build, maintaining four vendors’ stops making economic sense. We expect the category boundary to move accordingly.
4. Test data provisioning becomes the binding constraint
Authoring is largely solved. Maintenance is being solved. The next thing standing between a team and comprehensive coverage is the ability to establish the business conditions a test needs, particularly in ERP, where data is a consistent state rather than a row you can insert. We expect this to be the most-discussed unsolved problem in testing by 2028.
Hold Us To These The 2022 predictions are checkable because we wrote them down with a date attached. These four are stated the same way for the same reason. If two of the four are wrong by 2028, that is a fair thing to point at, and more useful than a forecast vague enough to always be right.
Where This Leaves Sofy
Practically, the platform today is five specialist agent lines, web, mobile, security, failure analysis and test authoring, plus enterprise coverage for SAP and Dynamics 365, coordinating through a shared context layer so a journey crossing several of them is validated as one thing.
The through-line from 2017 is narrower than the product surface suggests. It has always been the same bet: that the work of describing what should be true is worth a human’s time, and almost everything downstream of that is not. No-code was one attempt at acting on that. Agents are a better one. Whatever comes after agents will presumably make this look primitive too, which is the correct expectation to hold about your own current technology.
The Bottom Line
Between 2022 and 2026, testing stopped being about how tests get written and became about who decides what a test should do, and then, in ERP, about whether the software making decisions can be trusted to make them.
The no-code era answered a real question and the wrong one. Lowering the barrier to authoring never addressed the cost of ownership, which is where the money always was. Agents address it by removing the artefact rather than making it easier to produce.
What that opened up is a set of problems that did not previously exist, verifying agent judgment, provisioning data for tests nobody wrote, and observing correctness in systems that never error. Those are better problems than the ones they replaced. They are not smaller ones.