Carpe Data alternatives: ordering an investigation is not running one

Carpe (formerly Carpe Data) rebranded in July 2026 and made its flagship claims products agent-callable. The launch line is precise: a carrier's own AI systems can order an investigation without a human touching a screen. Ordering is not running. On Carpe's own Investigative Reports page, fulfillment of that order is scheduled at one to four days across three tiers, with human analyst verification.

This is not a takedown. Carpe is a decade-old claims intelligence vendor with real carrier distribution, and exposing investigation services over the Model Context Protocol is ahead of most of the market. The question for a carrier shortlisting Carpe Data alternatives is narrower than who is better. It is what work is still sitting on your side of the wire after you buy an agent-callable evidence surface, and who does it.

What follows: what Carpe sells in 2026, what the headless and MCP claim says literally, the five jobs a tool surface leaves on the carrier, what a state DOI exam asks the record to contain, the alternatives sorted by the gap they close, and a six-question evaluation checklist. It sits in the competitive cluster under our AI fraud platforms compared in 2026 buyer guide, alongside the same argument applied to the industry data utility in Verisk alternatives.

Carpe, known as Carpe Data until July 2026, is a claims intelligence vendor that sells open-web evidence on claimants, injury-lifecycle monitoring, human-verified investigative reports, image and document forensics, attorney-advertising signal, and an SIU case workspace. Founded in 2016, it is headquartered in Santa Barbara, California, with a second office in Lisbon, Portugal.

The rebrand was published on July 9, 2026 in a post titled Carpe Data is now Carpe, where the company describes itself as having operated for a decade and frames the name change as a move from data provider to claims intelligence suite. Max Drucker is founder, CEO and President, per the Carpe leadership page, with Geoff Andrews as COO, Andrew Arrastia as CFO, Scot Barton as CPO and Mike Nichols as CISO. The company raised a $6.6M Series A led by Aquiline Technology Growth, announced May 23, 2017.

Twelve days after the rebrand, on July 21, 2026, Carpe launched Carpe Claims with three new products. Carpe Case Management, per the release, "gives claims and SIU teams a single workspace where every party, vehicle, document, and piece of open-web intelligence on a claim lives in one place ..." Carpe Vision "brings forensic image analysis to claims: AI-generation detection, manipulation analysis, and metadata verification ..." Carpe AdWatch tracks attorney advertising and litigation marketing activity across markets, which is a signal nobody else has productized at this layer.

The claims line now runs Online Injury Alerts, Investigative Reports, Case Management, AdWatch, Vision, Recovery and LTC, with the Minerva Reasoning Engine, Carpe Cyber and Risk Control on the commercial side. The Case Management product page markets the workspace as an AI-native SIU cockpit and lists Ring Hunter for autonomous fraud-ring discovery, an AI Red-Flag Engine, AI Auto-Dial Interviews, Accident Reconstruction, a Red Team Adversarial Assessment and a One-Click Litigation Package, with a Claims Reasoning Engine underneath. Read as a whole, that is a product line covering evidence acquisition, monitoring, forensics, and a place to keep it all.

Scale is vendor-stated and unaudited, which is normal for this market and worth labeling anyway. Carpe states on its own site that it has processed more than 10 million claims since 2016, surfaced fraud evidence on more than 500,000 of them, and reclaimed more than $500 million for carriers, across 40-plus deployments. On distribution, the July 21 press release says Carpe is in production with a majority of the ten largest US property and casualty carriers, while the July 9 rebrand post says 40-plus carriers in production including nine of the top twenty. Two framings, twelve days apart, both from the vendor and neither independently audited. Carpe is also listed in the Duck Creek partner directory.

Model Context Protocol is a standard interface that lets an AI system call an external service directly instead of through a human-operated screen. In the July 2026 launch, Carpe states that its flagship products, Online Injury Alerts and Investigative Reports, now support fully headless operation, native MCP support, and complete agentic interactivity.

That is ahead of most of the market and it deserves to be said plainly before any contrast. Verisk shipped MCP connectors into Anthropic Claude on May 5, 2026, with two initial connectors covering underwriting and restoration data. Carpe exposed services rather than data. Exposing a lookup to an agent is a plumbing change; exposing an investigative service is a commercial one, because the agent is now placing an order that a supplier has to fulfill.

The release states the capability in one sentence: "A carrier's own AI systems can order an investigation, monitor a claimant, question the findings, and route the results without a human touching a screen." Drucker's line in the same release is the shorter version: "We spent ten years building the deepest view of claims activity on the open web. Now we've given it hands."

Read the sentence again for its verbs: order, monitor, question, route. Those are the four things an agent can now do to a Carpe service. Three are procurement actions and one is a query. None of them is deciding what a specific claim needs, running the pulls concurrently, reconciling results that disagree, judging when the file is complete, or writing the conclusion. That is not a criticism of the wording, which is accurate and carefully drafted. It is a description of where the boundary sits.

Two neutral notes for anyone mid-evaluation. As of this writing, no named reference customer has been published for Carpe Claims, which is unremarkable three weeks after launch and is still the kind of thing a procurement file eventually wants. And the agent-callable move is a category pattern rather than a Carpe quirk, so a carrier evaluating this in 2026 should expect the same surface from several vendors inside a year, and should decide now who orchestrates across them.

An MCP tool call is a procurement action. It places an order with a supplier who then fulfills it. On Carpe's own Investigative Reports page, that fulfillment is AI discovery plus human analyst verification, delivered on three tiers: Investigative Snapshot in about one day, Standard in two to three days, and Pro+ in three to four days.

The same page carries the only public pricing mechanic Carpe publishes: "No Hit, No Full Charge," a discounted rate stated as 50% when a report surfaces no substantive evidence. That is a fair structure and the mark of a well-run evidence service. It is also a per-report unit, which is the tell. A service priced and scheduled per report is a supplier relationship. An investigation is not a report. It is a set of them plus the judgment that ties them together, and the judgment is not on the price list.

The structural version of the problem shows up in the state numbers. In fiscal year 2023-24 the California Department of Insurance Fraud Division received 12,559 suspected fraudulent claims in its automobile program, assigned 602 new cases, made 272 arrests, referred 354 cases to prosecuting authorities, and put the potential loss at $207,629,944. Divide the first two figures and roughly one in twenty referrals became an assigned case. That derivation from the department's own numbers is what happens everywhere the capacity to raise leads outruns the capacity to work them, and it is the same shape inside a carrier as it is inside a state fraud bureau. The auto line is simply where it is measured and published.

Inside a carrier the same arithmetic runs on a shorter clock. Manual investigation of a flagged claim takes 14+ days per case and one investigator carries 200+ cases, so making it cheaper and faster to raise a flag or order an evidence pull moves the numerator without touching the denominator. An investigation agent moves the denominator. Hesper runs 15+ investigation phases in parallel on a flagged claim - document forensics, open-source research, statement cross-reference, timeline reconstruction, financial-pattern analysis - and returns a reviewable, audit-ready record in hours, not weeks. The phases run concurrently because per-case attention is not the constraint it is for a person working a queue.

A tool surface answers requests. An investigation decides what to request. Between the two sit five jobs that no evidence vendor performs on a carrier's behalf: selecting the evidence a specific claim needs, sequencing and parallelizing the pulls, reconciling results that contradict each other, judging when the record is complete, and writing the finding.

1. Selection. A soft-tissue bodily injury claim in a litigation-heavy venue needs a different evidence set than a total-loss theft claim on a two-week-old policy. Something has to decide which tool calls a given claim warrants before any of them fire, and getting that wrong costs money in both directions.
2. Sequencing. Evidence has dependencies. What the recorded statement says determines which records are worth pulling; what the medical file shows determines what the surveillance window should be. Running that serially is most of why a case takes 14+ days. Running it concurrently is a design decision, not a byproduct of faster tools.
3. Reconciliation. Evidence disagrees more often than it agrees. A monitoring alert, a metadata flag, a treating physician note and an adjuster's field observation routinely point in three directions, and the work is deciding which of them survives contact with the others.
4. Stopping. Every investigation has a point past which more retrieval stops changing the conclusion. Ordering reports beyond it burns budget and calendar. Stopping short of it produces a file that will not hold in an examination under oath or a coverage suit.
5. Writing the finding. The SIU lead, defense counsel and the state examiner all read the same artifact: a written conclusion with its basis. That is authorship, not retrieval, and it is the step that turns a pile of exhibits into an investigation.

Take a live file. AdWatch signals that the venue is saturated with attorney advertising for this injury type. An Online Injury Alert surfaces a video of the claimant carrying a kayak. Carpe Vision flags inconsistent metadata on a submitted damage photo. The medical file documents a restriction the video appears to contradict, and the treating provider has appeared on nine other files this year. Four signals, four sources, one claim, and not one of them is a conclusion. Something has to decide whether the video predates the loss, whether the metadata anomaly is a compression artifact or a manipulation, whether the venue signal is probative or merely prejudicial, and whether the whole thing supports a denial, a referral, or a payment with a documented note.

Reconciliation is where the market is weakest, and it is measurable. In Verisk research published in March 2026, based on 1,000 US consumers and 300 insurance claims professionals, 66% of insurers believe digital media fraud goes undetected often or very often, and only 32% are very confident identifying deepfakes. An ACFE and SAS anti-fraud survey reported in May 2026 found only 7% of anti-fraud professionals say their organization is more than moderately prepared to detect AI-charged fraud, with no insurance respondents reporting more than moderate confidence. Those are not retrieval failures. They are judgment failures under evidence load, and adding evidence per claim without adding judgment makes them more likely, not less.

This is the part that is easy to underprice in a demo, because retrieval demos well and judgment does not. Hesper is built at the other end: built-in fraud detection on the claim population plus the full investigation on each flag, which is what "from fraud detection to fraud resolution" means operationally. Detection is upstream; investigation is downstream. Carpe occupies part of the upstream and a large share of the evidence surface. The five jobs above are the layer underneath, and a carrier either staffs them, builds them, or buys something that runs them.

A state examiner does not ask what you retrieved. It asks what you concluded and on what basis. California's SIU regulations, 10 CCR 2698.36(a), list five things an investigation shall include, and only three of them are evidence work. The fifth is authorship: a written summary of the findings and the basis for them.

• (1) A thorough analysis of a claim file, application, or insurance transaction.
• (2) Identification and interviews of potential witnesses who may provide information on the accuracy of the claim or application.
• (3) Utilizing industry-recognized databases.
• (4) Preservation of documents and other evidence.
• (5) Writing a concise and complete summary of the investigation, including the investigator's findings regarding the suspected insurance fraud and the basis for their findings.

Three more provisions decide how this lands on a vendor decision. Under the California SIU regulations, 10 CCR 2698.41 lets the commissioner examine an insurer's SIU "including operations undertaken by entities under contract with the insurer," so routing the work through a supplier does not route the exposure with it. 10 CCR 2698.40(b)(8) requires the annual report to state the number of claims processed by the insurer and the number referred to the SIU, which puts a version of the referral-to-investigation ratio into a filed document. And 10 CCR 2698.42 sets penalties of up to $5,000 per act of non-compliance and up to $10,000 per willful act.

The pattern is national rather than Californian. NAIC Insurance Fraud Prevention Model Act #680, Section 11, states that "Insurers shall have antifraud initiatives reasonably calculated to detect, prosecute and prevent fraudulent insurance acts," and lists two ways to meet it: fraud investigators, who may be insurer employees or independent contractors, or an antifraud plan submitted to the commissioner. New York Insurance Law 409 requires qualifying insurers to maintain a full-time SIU separate from underwriting and claims, sets investigator qualification standards, and requires an annual report to the Superintendent by March 15 describing the insurer's experience, performance and cost effectiveness in implementing the plan.

Read together, the regulations are indifferent to who assembled the record and specific about what the record has to be. That is good news for automation and bad news for any output that is a bundle of retrieved artifacts with no authored conclusion attached. Hesper logs every step the agent takes with its sources, reasoning and timestamps, so the summary of findings and the trail underneath it are one artifact rather than two documents somebody reconciles a year later. The wider argument about what makes an AI-produced investigation defensible, including the NAIC AI Model Bulletin traceability point, is made in the defensibility standard for AI fraud investigation. This section is only about what the record has to contain.

Carpe Vision is image and document forensics for claims: AI-generation detection, manipulation analysis and metadata verification, per the launch release. Shipping it in 2026 is well timed, because the published market data says carriers are receiving manipulated submissions at volume and are not confident about catching them.

The Verisk research published March 17, 2026, fielded across 1,000 US consumers and 300 insurance claims professionals, puts numbers on both halves. 98% of insurers agree AI editing tools are fueling a rise in digital fraud and 76% say manipulated submissions have grown more sophisticated, while only 32% are very confident they can identify deepfakes. On the supply side, 36% of consumers said they would consider digitally altering a claim image or document, rising to 55% among Gen Z respondents. The demand for this product is not a vendor invention.

The practitioner version of this predates the survey data. Scott Clayton, Head of Fraud at Zurich UK, told the company's own broker publication in January 2024 that "There's software that is easy for fraudsters to access, which can manipulate images." Two years on the tooling is cheaper and better. Our own coverage of that shift, and of what a manipulated-media finding has to survive, sits in deepfake insurance claims and AI-altered evidence in 2026.

Give Carpe this section outright: forensic image analysis is real work, it is needed now, and a vendor with a decade of claims context is a sensible place for it to live. The distinction that matters to a buyer is scope. Document and image forensics is one of the 15+ phases Hesper runs on a flagged claim, not the product itself. A manipulation flag on a photo is an input to a determination in the same way a cross-carrier match is a lead and not a verdict, an argument made at length in the limits of cross-carrier fraud data networks. Whether the forensics come from Carpe, from Hesper, or from both, something still has to weigh that flag against the medical file, the recorded statement and the timeline.

The main Carpe Data alternatives are Skopenow for open-web and OSINT evidence, FRISS, Shift Technology and Verisk ClaimDirector for fraud detection and scoring, Clearspeed for voice-based intake triage, Command Investigations for outsourced field capacity, and Hesper AI for the investigation layer that resolves a flagged claim end to end. They are not substitutes for each other.

That is the whole difficulty with a Carpe shortlist: it mixes layers, which is how a carrier ends up comparing an evidence vendor to a scoring engine to a staffing firm and calling it a bake-off. Sort by the gap instead - open-web evidence, detection and scoring, intake triage, damage estimating, human investigation capacity, or the investigation layer itself. The table maps each name to the layer it occupies.

The nearest like-for-like to the original Carpe Data product is Skopenow, an OSINT platform for SIU investigators, adjusters and claims managers covering entity investigation, link analysis for fraud rings and claimant due diligence. It is a tool used inside an investigation, which is precisely what an investigation agent would itself commission.

Two options on the list are not Carpe substitutes at all but compete for the same budget line, which is worth saying out loud. Clearspeed raised a $60M Series D led by Align Private Capital in June 2025, taking total funding to $110M, and sells voice-based risk assessment at intake: it tells a carrier which claims to look at, before any evidence is pulled. Outsourced field investigation sells the same outcome as headcount. Command Investigations acquired the insurance division of CoventBridge in July 2026, giving it roughly 900 investigators across all 50 states, with the acquired division serving more than 200 carriers, TPAs and self-insureds. That is real capacity, and surveillance and examinations under oath still need people. It also scales linearly with hiring and runs on the same per-case clock as the SIU it supplements.

On the detection side, FRISS scores at first notice of loss and through the lifecycle with an interface SIU teams are comfortable in, Shift Technology pairs detection with handler-assist agentic AI for claim handlers, and Verisk supplies contributory data with scoring on top. All three raise flags; none of them works the flag to a conclusion. The longer treatment of that split, with the ISO ClaimSearch scale numbers, is in Verisk alternatives.

Hesper sits on that list at one layer and only one. It ships built-in fraud detection plus end-to-end investigation on each flagged claim, and it works standalone rather than assuming a detection contract is already in place. It does not do damage estimating, it does not run a contributory data pool, and it does not replace a decade of open-web collection. It can consume the same evidence surface a carrier already pays for and decide what to do with what comes back.

The practical way to price the difference is to list what an investigation actually requires and mark who supplies each item. An agent-callable tool surface supplies retrieval and monitoring. A carrier can build orchestration on top of it. An investigation agent supplies both, plus the judgment steps in between.

Read the middle column carefully during an evaluation, because it is the one that never appears in a demo and always appears in the implementation plan. "Carrier must build" means an orchestration project: policy and prompt design, an evidence schema, conflict handling, stopping rules, audit logging, and a review workflow for the SIU lead. Some carriers should build it and have the platform team to do it. Many discover it on the plan after signing for the tool surface, which is a common failure mode in agent-era procurement.

For a CIO the question is governance rather than capability. If a carrier's own agent is placing tool calls to three vendors, the reasoning that connected those calls lives in whatever orchestrated them. If that is a general-purpose assistant, the record a DOI examiner eventually reads is a chat transcript. If it is a purpose-built investigation agent, the record is the case file, with sources and timestamps attached to each step. Where the reasoning is logged is not an implementation detail. It decides what the carrier can produce two years later when one file gets pulled.

The cost of an evidence-only stack is not the invoice. It is the widening gap between flags raised and flags resolved. Referral volume is cheap to grow and investigation capacity is not, so a cheaper and faster evidence surface increases the number of claims that deserve work without increasing the number that get it.

Start from the baseline. The Coalition Against Insurance Fraud puts US insurance fraud at $308.6 billion a year, with fraud occurring in about 10% of property-casualty losses. Against that, a manual SIU fully investigates roughly 25% of the claims it flags, because each case runs 14+ days and one investigator carries 200+ cases. The other three quarters are paid, denied without full work, or queued indefinitely. Cost per manually investigated case sits near $2,500 against roughly $150 with automated investigation, and coverage moves from about 25% to 100%. Those last figures are Hesper internal benchmarks, and the point of them is the ratio rather than the decimal.

The clock the investigation runs against is not the SIU calendar. It is the policyholder's. The J.D. Power 2026 US Property Claims Satisfaction Study, fielded from December 2024 to December 2025 across 5,093 homeowners claimants, put average repair completion at 29.6 days and average time to final payment at 40.7 days, both improved year over year. A 14-day investigation inside a 40-day payment cycle is a third of the file. A one-to-four-day evidence order inside that same cycle is better, and it is still a serial step in a process the claimant experiences as waiting.

For the person holding the budget, the choice usually gets framed as renewing an evidence contract, adding investigators, or buying the layer. The first two scale linearly with spend, and the California referral-to-case ratio is what linear scaling looks like at state level after decades of it. The only variable that moves coverage from a quarter of flags to all of them is removing the per-case human hour as the unit of production. That is what makes the investigation layer a loss-cost lever and not an SIU efficiency project, and it is why this decision belongs in a loss-ratio conversation instead of a tooling one.

Both can be true at once. Carpe is an evidence and monitoring surface with a workspace on top of it. Hesper is the layer that decides what evidence a claim needs and turns what comes back into a finding. A carrier can run both, and the modal stack for a large carrier probably will. What decides the shortlist is six questions, asked of every vendor in identical words.

1. Who decides which evidence a specific claim needs? If the answer is the caller, you are buying a tool surface and you own selection.
2. Who reconciles evidence that contradicts other evidence, and where is that reasoning written down afterwards?
3. Who authors the summary of findings and its basis, as 10 CCR 2698.36(a)(5) requires, and in what format does it arrive?
4. What is the turnaround per investigated claim, not per report? Per-report SLAs are supplier metrics; per-claim cycle time is what shows up in reserving and cycle-time numbers.
5. What does a DOI examiner see when they pull one file two years from now? Ask for a redacted real example, not a dashboard screenshot.
6. Does the product work standalone, or does it assume an orchestration layer you have to build? If the second, price the build before comparing the licenses.

Hesper answers those six by design, not by configuration: built-in detection so it runs standalone, 15+ investigation phases in parallel per claim, an audit trail logged with sources, reasoning and timestamps that ships as the deliverable instead of a separate compliance document, and hours, not weeks, per case. It complements detection and data vendors instead of replacing them, and it can consume the same evidence surface a carrier already buys. The objective is not a shorter vendor list. It is to make every flagged claim investigable, which is the only version of this that moves the loss number.