The Last Software You’ll Buy

Part 1. The Old Era

I have written 4 times this year about what is changing in insurance. Why building your own software is a bias rather than a strategy. Why orchestration is the bottleneck. What will happen to the claims handlers. And why AI, as it has  evolved  over the last 4 years, is becoming a new category of software.

This one is the prequel. It is about why all of it is happening, and why it took 50 years to get here.

I have built and delivered software products for 16 years. Long enough to notice that every one of them fell into one of 5 categories.

• Creates work that didn't exist.
  The first commercial catastrophe model shipped in 1987. Hurricane Andrew proved it in 1992. Cat modeling became standard practice within ~8 years.
• Automates existing work.
  MetLife installed a UNIVAC in 1954, the first large-scale computer in life insurance. It was bought because the company could not hire enough clerical staff to process its own paperwork.
• Augments existing work.
  Guidewire launched in 2001. It took roughly 15 years to become the default choice for carriers needing a core system.
• Transfers existing work to someone else.
  Self-service FNOL portals arrived in the late 1990s, moving to mobile phones in 2000s. It took about 20 years to deploy them seriously. The work finally moved to the policyholder.
• Eliminates the need for the work.
  Water sensors stop the leak before it becomes a claim. In active deployment for over a decade. Still not normal.

5 categories. 5 different business cases. Same 8 to 20 years of normalization.

The adoption time did not change with the technology. That should give you a hint why the Insurance industry has notorious reputation of being slow.

In 2026, I can build an enterprise-grade product in 6 months or less. The industry still needs decades to decide if they want one.

A few problems cause it, some are interconnected.

Insurance runs on functions. Each function is a cluster of tasks. For 50 years the industry bought software cluster by cluster. Zylo's 2025 index puts a large enterprise at around 660 applications, and 56% of software purchases happen outside IT entirely. Nobody decided that. It accumulated. Every few years the company runs another transformation. Every new CIO starts another consolidation. Software accumulation doesn’t stop.

Half the IT budget goes to keeping that pile alive. West Monroe surveyed 300 insurance executives in October 2025. 54% spend more than half their IT budget maintaining what they already run. Roughly half the licenses they pay for sit unused. There is very little money free to bring in something new, while the function still has to adapt to a new world reality every year. That adapting gets absorbed by employees working harder.

Insurance companies still try to build their own software, acting like technology companies. There are over 30,000 software companies in the world and insurance decided the answer was 30,001. I have written a long piece on why that decision keeps getting made against the evidence, so I will keep this short. The build is not the expensive part. Maintaining and continuously improving it for 20 years is. It almost always fails or ends in paralysis, which is harder to see and harder to fix.

Nobody gets fired for waiting. Being early on a failed project is career-ending and personally attributable. I have seen leaders who refuse to bring in transformative products because it’s too risky for an end of year bonus. Being 5 years late is invisible and collectively attributable. I have also seen people make huge career moves by “fixing fires”. We built an industry where the safest career move is to do nothing, and then we wonder why nothing happens.

The right software matters. Buying or building software is not enough to bankrupt you on its own, but enough to push you toward the two things that do. A bloated estate raises your cost per policy. From there you either price to compete and under-reserve, or you hold price and go subscale until somebody buys you. Deficient reserves are the leading documented cause of insurer insolvency. Subscale is the leading reason carriers get acquired.

Software never appears on the death certificate. It appears in the Combined Ratio first.

Year over year we build and buy software to change the work itself. Create new work, automate it, augment it, transfer it, eliminate the need for it. The cycle is still running. The era that produced it is already over and giving way to AI. It will take 8 to 15 years for anyone to notice, same as everything else.

Part 2. A New Kind Of Software

For the first time, one piece of software can do all 5 categories. Automate, augment, create, transfer, eliminate. Same system, same day. It feels like magic. Our demos of the AI Agent still puzzle many.

That has never happened before. Every product I listed earlier could do one.

The impact is not visible yet, because it arrives in phases. Here is how I see it.

Phase 1: Every category becomes an AI version of itself

The noise is almost unbearable. Every software company is adding AI, at minimum to their name.

In March 2024 the SEC had to invent an enforcement category for it. They call it AI washing. Delphia paid $225,000. Global Predictions paid $175,000. If you ever wondered whether lying about AI pays, the market has now run the experiment and published the price. Making fancy slides, unfortunately, still works.

Gartner expects 40% of enterprise applications to ship task-specific AI agents by the end of 2026, up from under 5% in 2025. Roughly 11% of agentic AI use cases reached production last year. Everyone is shipping agents. About 1 in 9 survives contact with a real operation. I never liked the Pilots, and there is a graveyard of them the size of Manhattan.

Everyone is working out how to apply it. The bigger companies just buy an AI company, because that is cheaper than rebuilding your architecture and it lands as a press release the same week. Everyone mastered Product Launch videos to increase interest and valuation of the business.

You are not buying capability. You are buying a slide.

Every company I know either bought someone, pivoted, or added an AI feature. More AI assistance inside your software, please. Insurance is not an exception. 9 out of 10 conversations I have are Build vs Buy.

We are back to the time when everyone built their own database and ran their own email server. If coffee makes your employees more productive, it doesn't mean you should start a coffee farm. Be the company you intended to be, unless it is a real hard pivot, and for most of you it is not.

A lot of budget is being spent on a problem that was never real. And most of the people spending it will have moved on before anyone checks. If you fire your Technology Executive, who will deal with all the mess?

Phase 2: The categories blend

I have noticed something no one talks about, at least not openly.

Augmentation used to mean a human did most of the work. The orchestration, the execution, all of it. Automation was brought in to make that human faster, cheaper, and less wrong.

An AI Agent can now do both. The Agent does most of the work. The human reviews, corrects, gives feedback, oversees.

That is a different operating model, not a better tool. The Agent becomes the center of the work. The human becomes the center of the cognitive power.

Most leaders are still buying AI Agents to do automation when they should be transferring the work to them. Automation and Augmentation stopped being two categories. They collapsed into one, and the business case is still written in the old language.

The market can feel it even where it cannot name it. Core platforms are buying AI layers. AI companies are adding core capabilities. Duck Creek bought Send Technology in July 2026, an AI-native underwriting orchestration engine, and folded it straight into their Agentic AI Platform. Sapiens bought two companies in one week in April 2025.

Every one of these deals gets announced as the first and only platform of its kind. They cannot all be first. What they are is scared, and correctly so. We are actively in conversation with core product companies to be their AI Layer. They know, we know. Insurance companies stay confused.

Your vendors are deciding what they want to be. You are watching it from the outside.

The companies that understand the inversion first will not be competing with you on cost. They will be doing things you cannot do at all. Your competitor will drive Ferrari, and you will drive old Fiat. Hey, nothing wrong with Fiat, but everyone wants Ferrari.

Phase 3: One category of software

Eventually there is just Agentic AI, or whatever it ends up being called. The frontier labs are spending enormous money on investment, collaboration and acquisition to convince the world that their software is the only software you need. The US and China compete for dominance. Everyone else looks for a spot, preferring not to see what is coming.

Watch what Salesforce did.

First, they would build their own model. Then Agentforce. Then, on 26 August 2026, Claudeforce with Anthropic, and their CRM went inside somebody else's AI. The product is called Salesforce in Claude. 37 prebuilt sales skills. Salesforce's own announcement says the user interface is the AI.

On the same earnings call, Benioff said: "This nonsense of the SaaSpocalypse, I think it's time for it to stop."

He announced it and denied it in the same hour. That is not hypocrisy. That is a CEO who can read the trend line and still has a quarter to get through.

Now apply that to your own estate. If the biggest software company on earth accepts that nobody logs into their product anymore, and an Agent does it for them, then every application you own becomes a place where data sits and rules live. Something else does the work on top of it. Satya Nadella called it 2 years ago and it was lost in the void of AI noise.

Your 660 applications stop being 660 products. They become 660 connectors.

2 years ago that was science fiction. It is now happening at the top of the market, and the top of the market is where this always starts.

I have written separately about where this lands for insurance specifically. Core systems demoted to databases. Agents doing the work on top of them. People governing it rather than operating it. That is the AI Operating System, and the interesting part is that carriers are pulling it into being one feature request at a time, not vendors imposing it. Top of market adapting, mid-market sees it as signal to Build your own AI, bottom of the market has no time for anything, but to continue survival battle.

Phase 4: Doomsday

One superintelligence, maybe several models with different jobs, running everything. One system on your phone. Or your device. Or closer than that. No software to think about at all. If it happens, 20 to 50 years. But I have watched the last few years, and I would not be shocked if ai-2027.com plays out roughly as written.

Either way, almost nobody in this industry is planning for it. So, it will arrive the way everything else arrives here. Late, and as a surprise.

Here is the part that should bother you.

Phase 1 started two years ago. Phase 2 is happening now, in your vendor list, whether or not anyone briefed you on it. Phase 3 is visible at the top of the market. Phase 4 is preached by LLM models and once in a while gets slowed down by another Government request.

And most insurance companies are still deciding whether to build it themselves.

Part 3. Claims, The Proof

Everything above is theory until it touches real work. So, take the function where insurance spends the most money and makes the most enemies.

Insurance is a $7.1 trillion industry. Roughly 70 cents of every premium dollar goes to paying claims and adjusting them. Another 29 cents runs the business. In a normal year that leaves about a cent of underwriting profit, and the industry survives on its investment income. When things go sideways for a few years in a row, companies get acquired or they die.

We had customers who brought in AI to fix their Combined Ratio too late and went out of business six months after signing with us. Good run. But an AI Agent cannot be your last resort. If Insurance companies were a bit more open sharing the real state of business, we would be able to help better.

Insurance is a promise. A claim is when your company gets tested on it, and most of the time it fails the test. Accenture found more than 30% of customers were not fully satisfied, to say the least, with a recent claims experience. Of that group, 30% had already switched and 47% were thinking about it. They put up to $34 billion of premium at risk every year.

That is the price of being bad at the only thing your customer will ever judge you on.

You cannot wait 45 days on a broken car window. You cannot wait 2 weeks to find out a document was missing the whole time.

The reason is that a claim looks simple and is not. Submit with evidence, get a decision, get back to normal. Underneath that are dozens of interconnected activities, multiple stakeholders with different agendas, and over 30 tasks that all have to be done correctly to settle one ordinary claim.

The queue is long. There are fewer people than there were, and there will be fewer still in ten years. The Institutes and BLS estimate more than 400,000 unfilled insurance positions as roughly half the industry retires by 2035. The younger generation does not consider claims management a viable career, and they are not wrong. Why become a claims expert if AI can already do most of it.

What the job actually is

I have described a claims handler's day in full detail elsewhere. The short version.

The work is notoriously boring. The genuinely complicated and interesting things happen maybe once or twice in the entire process.

Most of the day goes into claim intake. Reading, classifying, extracting. Then moving that data between systems. Chasing people for missing information. Uploading documents. Hunting for the password to an archive file. Reading evidence, then reading it again. Explaining coverage to someone who never read the policy. Calling service providers for quotes. It goes on.

The real value of a person arrives in one moment. Who is liable, and what outcome should we reach. Everything else is supporting work, and most of it is admin.

What you are paying for is experience. The ability to recognize a situation and apply judgment to the case where the logical answer is A and something tells you it is B. AI can get there too, with enough running, training and access to historical data.

Cluster those tasks and you get 7 milestones across 3 stages. That is the end-to-end lifecycle.

Claim Intake. Capture the loss at first notice. Collect evidence and chase missing information. Verify policy coverage. Categorize and triage. Assign to handler or team. Identify large losses. Identify complaints. Register the new claim. Update an existing claim. Create the claim summary. Communicate with the policyholder.

Claim Processing. Assess liabilities. Assess damage and repair cost. Calculate reserve. Identify and pursue recourse. Generate handler action items. Run leakage checks. Mitigate fraud. Make the decision or recommendation. Communicate with service providers and policyholder.

Claim Closure. Handle payment. Run subrogation. Run recoveries. Archive the case file. Run the quality check. Run the compliance check. Close the claim. Communicate with the policyholder.

29+ of them. Read that list again and count how many need a human being.

What an AI Agent for Insurance Claims Does

This is where we deliver value across 4 of the 5 categories from the beginning of this article. Automate, augment, transfer, create. We do not eliminate the need for the work. Not yet.

When we deploy an AI Agent for Claims, we go after one or all four:

•      Automate the tasks a human does not need to do. First Notice of Loss. Collecting missing information. Last time I counted 17-20 different tasks.

•      Augment the human. Give them a decision-ready claim. The Agent reviews it and presents its assessment in a digestible form, so the person says yes or no. Some of our customers call it 1-click claims, just like Buy in 1 Click on Amazon. When I explain it to insurance leaders, they don’t believe me. I understand.

•      Leave alone what should not be touched, at least for now. Rejections. Fraud Investigations.

•      Create new work. Reviewing the Agent's output. Fine-tuning its configuration. Acting on the insights it produces about your own operation. We often help our customers create new roles, some claim handlers become Agentic Claim Handlers. They actually LOVE their job now.

Our customers report some great results, but they are all different, so make it harder to clearly articulate the value of an AI Agent beyond it will be FASTER, CHEAER, BETTER. I know what you are thinking.

So, ask yourself one question. Do you need 20 different pieces of software to handle a single claim, or can you let an Agent do it?

The Agent connects to your software and uses it instead of a human, and the human connects to the Agent. That is the simplest possible version of a new operating model. It is also the version where your existing software ecosystem stops mattering very much. You don’t need to replace or migrate anything.

Simple to describe, hard to build. Agents on their own are not enough: you need state that survives a 3-week wait on a medical report, defined behavior when a step fails, compensation logic when step 5 breaks after steps 1 through 4 succeeded, and an audit trail a regulator will accept. That is orchestration, and I have argued it is the real unsolved problem in insurance technology. The agents are extraordinary musicians. They still need a conductor, at least for the next few years.

Why the business case is always too small

Almost every conversation in this industry is about automation. When it comes to AI, the 2 words have become synonymous.

The vendors are not raising anything else either, because automation is easier to sell and easier to sign off. Everybody gets a clean business case, and nobody has to think.

So, most business cases get built the same way. Cost saved, hours saved, STP rate.

I recommend 4 buckets instead: money saved, spending avoided, revenue saved, revenue generated. That framing forces you past automation, because the last 2 have no automation answer.

If I give you back 50% of your claim handling time, you do not automatically get more claims. Claims efficiency is not a revenue-generating activity. Not until you start thinking about what that capacity does for marketing, sales, and everything else in the business.

The question nobody can answer

I ask executives what they would do with 50 to 70% of their team's time back.

Almost none of them have an answer.

Feels like criticism, but it just requires more thinking than a Friday strategy session. You would want a year to work out what it means for your company, your market, the economy, and society. But most of them will not spend an hour on it, because there is no line in the budget for questions that do not have a number attached.

If an AI Agent can do everything a human does, your problem stops being how much money you can save. It becomes what operations you need to hit your business objectives.

And then, hold on. What should your business objectives even be if an Agent does 80% of the work? What kind of business could you build? Is this still a pure P&L world, or do we finally get to talk about purpose?

Part 4. Why It Isn't Happening

I have sat in the room where this decision gets made. 5 reasons, and only 4 get attention.

Trust and accountability

Here is the case that ends every conversation about autonomous claims decisions.

In November 2023 a class action was filed in Minnesota against UnitedHealth over nH Predict, an algorithm used to assess post-acute care. The complaint alleges roughly a 90% error rate, measured by the share of appealed denials later reversed. It also alleges the algorithm stayed in production because only around 0.2% of policyholders appeal. In March 2026 a judge ordered disclosure of how it was built.

If your error rate only surfaces through appeals, and almost nobody appeals, then error rate stops being an operational input. It becomes something you find out about in discovery.

When a claims director raises this, they are not being difficult. They are being correct, and it is my job to have an answer.

The uncomfortable part is that insurance did this to itself 30 years ago, with rules instead of models. Colossus valued bodily injury claims at roughly a quarter of the top 100 US carriers. Regulators eventually made one carrier review around 2 million of them. One settlement term is the tell: the company had to stop requiring adjusters to settle on the software's recommendation alone.

A regulator had to formally instruct an insurance company that a human being was permitted to disagree with a machine. That was 2010. We adopted the procedure and skipped the lesson, which is that a system nobody can override is not decision support. It is a decision.

Regulation

EU AI Act, EIOPA's August 2025 opinion, GDPR Article 22, the SCHUFA ruling, FCA Consumer Duty, the NAIC bulletin in 24+ states, Colorado SB21-169, New York DFS Circular 7. I have read all of them, because we have to build to them.

6 regimes, 1 shape. The accountability stays with the carrier. It does not transfer to us and it will not transfer to anyone else. I asked that question in front of European regulators in the big official auditorium in 2025 and got no answer. “Who is to blame?” We don’t know yet, but let’s write more compliance rules. My procurement process just got 2 months longer.

Any vendor who tells you otherwise has never been through a regulatory review.

Data and systems

Legacy cores, inconsistent data, integration debt. Real, boring, and over-cited, because it is the only constraint you can name without implicating anyone at the table.

Labor agreements

In the Netherlands, works council consent under Articles 25 and 27 of the Works Councils Act covers introducing this kind of automation, and the sector collective agreement carries a digitization agenda and a redeployment mechanism. Nordic co-determination works the same way.

Almost nothing written about claims automation mentions this, which is a reliable signal about where the author has actually deployed. We have. It adds months, and it is legitimate.

Imagination

Never on a risk register. Always the actual blocker.

It is genuinely insane to me how much time and energy we spend arguing carriers into decisions that pay for themselves. They go build it in-house. 2 years later they are back with the budget spent, a Combined Ratio over 100, and customers telling them the experience is terrible.

I used to think this was an information problem. It is not. I would bring the logic and watch it bounce.

Here is what took me years to understand. The decision to build in-house is not a technical assessment. It is a rational response to a payoff structure.

Being early on a failed project is career-ending and personally attributable. Being 5 years late is invisible and collectively attributable. Building in-house is a 2-year decision that fails slowly, so the person who made it has usually moved on before the failure is legible.

The executive is not being irrational. They are optimizing a different function than the one you think they are optimizing.

I am not above this. I worked on Wall Street and inside Fortune 500 companies. You can run a long and successful career there without ever being right about anything. You just have to be wrong at the same time as everybody else. I switched sides to be a vendor, but politics didn’t change.

Show me the incentive and I will show you the roadmap.

Part 5. What Happens Next

This has happened before in insurance claims, with different technology.

• 15 years ago. Paper files. First notice of loss by telephone. Physical inspection as default.
• 10 years ago. Digital FNOL and self-service portals. Core platforms migration to Guidewire, Duck Creek, Sapiens, Keylane. OCR and RPA as the first automation wave. Claims administration moving offshore.
• 5 years ago. Remote adjusting, normalized after 2020 and never reversed. Photo-based estimating. Straight-through processing on simple lines. Telematics feeding motor.

Every wave removed tasks and added others. Every wave arrived with predictions that the adjuster would disappear, and counter-predictions that nothing would really change. Both were wrong, every time.

What kind of prediction can we make based on  history, just like Ray Dalio suggested.

First. The Agent gets adopted as capacity, not as strategy

Claims handling stops being a career anyone chooses, and carriers are forced to absorb the gap with technology because there is almost nobody left to hire. Plenty of companies will lag and go out of business on pure lack of workforce and inability to stay competitive. We already see modern carriers doing everything 10x better than the old ones. There is going to be a sea of dead bodies and a lot of business absorbed by the companies still standing.

This is not a new pattern. MetLife bought a UNIVAC in 1954 because it could not hire enough clerks. 2 computers were installed in the entire insurance industry that year. More than 20 the year after. The fastest technology adoption in the history of this business was driven by a staffing shortage, not by a business case.

That is what is coming. Not a strategy. A shortage.

No surprise that Tier 1 enterprises are partnering with the LLM companies. That is their attempt to reach Phase 3 through Phase 2.

Second. Humans handle exceptions and nothing else

The claim queue runs on the Agent. A person gets involved when the Agent is stuck, and resolving that case teaches the system, so the case does not repeat. Tesla ran billions of miles of driving data before shipping anything close to autonomy, and the interesting miles were always the edge cases. You need enough of them to optimize the majority of the workload.

Here is the part people miss. The exceptions are the hard claims. All of them. Every routine file is gone from the queue. What is left is contested liability, serious injury, fraud, and the claimant who has lost everything. How do you handle a claim when a car crashes into a plane on the airport runway. I heard a story about it. They are still figuring out what to do. That is a harder job than the one your team does today, done by fewer and more senior people, with an audit trail behind every decision.

Nobody is training for that job. Nobody is hiring for it. Nobody has written the competency framework. That is the argument I made in The Last Claims Handler, including what it does to the training pipeline when the entry-level work disappears first.

I do not expect anything extreme in the next 5 years, but 10-20 years is unknown. Most companies will keep lagging and deal with the consequences of deciding to build software instead of working out a business strategy and a new operating model.

Unless regulators pull a harder brake than the EU AI Act and the NAIC, insurance is going to change beyond recognition.

What I am not predicting

I am not predicting that claims handlers disappear. Everyone who has said that has been wrong. I am predicting the job gets smaller, harder and more senior, and that the carriers who plan for that will buy the ones who did not.

What To Do About It

Rethink the work. Stop trying to build your own software. It adds no value and it only feeds the mess in Phase 2. If you want to win, operate like it is already Phase 3, and hope we never get to Phase 4.

Do not buy 20 pieces of software to handle one claim. Do not build the twenty-first yourself.

Adopt the AI Agent and change the way the work is done.