Keep the system of record. Get rid of the people you hired to bend it around your business.

That is a narrower claim than the headlines about AI eating software, and it’s the one we’ll actually defend. We’re not arguing that SaaS is dead. The ledger underneath your operation - the accounting rules, the inventory math, the order states - earns its keep, and you should pay for it. What’s over is the multi-year army of customization consultants who used to sit between that ledger and your day-to-day operations. AI, paired with the Model Context Protocol, an open standard for wiring AI into the systems where your data already lives, now does the fit-and-integration job that was once a small firm’s entire revenue line.

Tallyfy sits at the layer this shift is moving toward, so weigh that bias as you read. The more we sat with the pattern, the narrower the claim got: the software does not lose, the human scaffolding around it does.

Buy the ledger, skip the rebuild

The buy decision was never the expensive part. The license for the system of record - the SAP, Oracle, NetSuite, or Workday underneath your operation - is a known number on a quote. What hurt always landed after the signature: discovery workshops, configuration, custom objects, the integrations, and the firm that billed every hour of it. We wrote a whole piece on the bill the brochure hides for heavy platforms, and the mechanic is the same here. That post-signature effort is what AI now eats.

The mood has data behind it. Retool’s 2026 build-versus-buy report, drawn from 817 builders, found 35% of teams have already replaced at least one SaaS tool with something they built, and 78% expect to build more internal tools this year. Read that carefully. Builders aren’t turning their backs on software. They’re turning their backs on the bill for a person to assemble it, now that the assembly got cheap.

Software was never the line item that wrecked the budget.

The mechanic underneath that number is plain. An AI agent can read the schema of your system of record and the API of the tool you want to connect, then draft the mapping between them without a person staring at field names for a week. The translation that used to be billable expertise is now closer to a prompt. So the gap between a generic platform and your business, the gap that justified the whole engagement, is the piece of the job that fell hardest in price.

What the consultant army actually sold

Strip an engagement down and the customization army sold three things: fit, plumbing, and translation. Fit meant bending a generic system to match how your business really runs. Plumbing meant wiring it to the other dozen tools in your stack. Translation meant turning a sentence like “approvals under five thousand dollars skip the regional manager” into configuration somebody could maintain after the consultants left.

All three are now things a capable AI does against a defined workflow, in hours rather than quarters.

Actually, that’s too clean, and we should say so. Most of those consultants did real work, and the best of them understood the business better than its own org chart did. They weren’t frauds. The delivery model they sold, humans hand-fitting code across many months, just stopped being the cheapest route to fit, plumbing, and translation. When the price of a capability falls through the floor, the people whose income depended on it staying high notice before anyone else.

Does your implementation consultant still earn the fee?

A question that keeps coming up in our conversations: when the integration work gets cheap, who in the room argues hardest against trying the cheap way? Often it’s the outside advisor whose fee depends on the slow way.

In one of those conversations, an operations lead at a manufacturer we work with described the pattern exactly. Their longtime consultant kept finding reasons not to test the AI approach: the data was not ready, then the risk was too high, then the timing was off, on and on. Every objection sounded prudent, and not one survived a small pilot. What caught us off guard was how reasonable the resistance looked from outside: the advisor had a billing model to protect, and that model only pays when a project runs long and manual.

Is that always self-interest? No. Sometimes caution is just caution, and a botched AI pilot can burn real trust. None of this argues against ever hiring help. A good process consultant earns the fee, and the discipline of process consulting survives AI without much trouble, because the hard part there is judgment, not configuration. The narrower target is the implementation army whose product was hand-built configuration.

That said, you can test which kind you are dealing with in an afternoon. Ask the advisor to scope the smallest possible AI-run version of your next process change, then watch what happens. The ones who shrink it to something testable are with you. The ones who let scope creep back in until the project needs them are defending the army.

There is a cleaner rule, and it is arithmetic.

Take the implementation statement of work and split it into two columns: data the system already holds, and judgment only your people have. The first column is what AI and MCP now handle for a fraction of the quote, the schema reading, the field mapping, the report wiring, the connector upkeep. The second column, deciding which approvals matter and where a human signs off, is the slice still worth paying for, and it usually turns out to be a sliver of the total. If the quote is ninety percent column one, you’re paying for labor that just got commoditized.

Pay for the judgment. Skip the keystrokes.

The ERP becomes a database with accounting rules

One line from that operations lead stuck with us: with AI handling the integration, the ERP becomes a glorified database with accounting rules. Strip away the bolted-on customization and that’s what’s left, and that’s fine. A clean ledger with solid rules is worth good money. It was always the scaffolding around it, the janky custom screens nobody liked, the brittle point-to-point connectors, the reports only one contractor knew how to change, that drained the budget.

The ledger was never the problem.

This is the same shift hammering the integration platforms. For years, wiring System A to System B meant a connector, a field mapping, and someone on call when an API changed under you. That whole category was middleware, and it billed like a tollbooth on traffic you generated. MCP and AI agents are turning that tollbooth into a sentence you type. You describe the data flow you want, the agent writes and runs the connection against a workflow you have defined, and the kludge of connectors you used to babysit just thins out.

The pricing trend pushes the same direction. Zylo’s index, reported by CIO, pegs average enterprise SaaS spend at $55.7 million across about 305 apps, flat in app count but climbing in cost. The pricing model is shifting too, from per-seat licenses toward usage-based bills. You’re paying more for the same shelf of tools, so the contrast worth wanting is a tool whose price sits on a public page, with no implementation contract hiding the real number behind it.

Where the work actually happens now

Clayton Christensen gave this its name decades ago. He called it the law of conservation of attractive profits: when one rung of a value chain turns modular and cheap, the money does not evaporate, it migrates to the adjacent rung that is still hard. The system of record is becoming a commodity. The customization labor wrapped around it is getting cheap faster. So the hard part has to land somewhere.

It lands at the layer where people and AI run the actual work - who does what, in what order, with which handoffs, and what an AI agent is allowed to touch.

That’s the layer we’ve spent more than ten years building, so for what it’s worth, treat our enthusiasm with suspicion. The logic holds without us, though. Hand an AI agent a job with no written process and it improvises, and improvisation doesn’t scale past the demo. Most companies have never recorded their processes in a form anything, person or model, can follow, which is why the orchestration layer is mostly greenfield. The biggest lesson we have learned building Tallyfy is that this layer holds the durable value, because it encodes your judgment instead of your vendor’s defaults.

So buy the ledger. Let the army go. Then spend what you save writing down the processes that people and AI will run on top of it, because that’s the labor that no longer outsources cleanly, and the work that decides whether your AI does anything useful at all.