Newsletter
Join the Community
Subscribe to our newsletter for the latest news and updates
AI underwriting tools promise more consistent pricing and better loss ratio management. Here's what actually works, for which carriers, at what cost.
2026/04/18
Last reviewed 2026/06/06
Picture the submission stack on a Monday morning at a mid-size commercial P&C carrier: 200 new accounts, ranging from small restaurants to light manufacturing to contractors. Each underwriter has their own pricing intuition, their own read on loss history, and their own interpretation of the appetite guidelines. The result is pricing consistency that varies by who is assigned the file — and a segment loss ratio that has been drifting upward for six quarters. The board wants to understand why. The underwriting chief does not have a clean answer.
This is the problem AI underwriting tools are designed to address. Not to replace underwriters — the most credible vendors in this space are explicit about that — but to give underwriters a consistent data foundation and to surface the signals that humans miss at volume. When it works, the loss ratio drift becomes measurable. When it does not work, you have spent significant budget on a model that does not generalize to your book.
The category label "AI underwriting" covers a spectrum of genuinely different capabilities that are worth distinguishing before evaluating vendors.
At one end are data enrichment and scoring tools: platforms that gather third-party data about a submitted risk — business type, revenue, claims history from external sources, safety indicators — and return a score or a set of signals for the underwriter to consider. Planck is the clearest example. These tools do not require the carrier's own historical loss data to function, though they improve with it. They reduce the manual research burden on underwriters and surface information that would otherwise require multiple data pulls.
In the middle are loss prediction models: tools that use the carrier's own historical data, combined with external signals, to predict the probability that a given risk will generate a loss within a defined time window. Gradient AI is the most cited example. These models require substantial training data and are most defensible in commercial lines segments where the carrier has meaningful volume.
At the other end are portfolio management and workflow routing tools: platforms that connect underwriting decisions to portfolio-level outcomes, managing which submissions get prioritized, which underwriters handle which segments, and how the aggregate book is tracking against target loss ratio. Federato sits here.
Cross-cutting these categories is a newer capability: generative AI for underwriting documentation — tools that read submission documents, summarize risk characteristics, and draft underwriting rationale. Sixfold specializes in this. This capability is distinct from predictive modeling; it does not produce a risk score, but it can dramatically reduce the time an underwriter spends on document review per submission.
For background on the underlying concepts, see our glossary entries on predictive underwriting and explainable AI.
The carriers most actively deploying AI underwriting tools are commercial lines carriers writing small-to-mid-market accounts, MGAs with concentrated expertise in specific segments, and specialty carriers where the data signals are distinctive enough to support model training. Personal lines carriers have been using predictive models for longer — the homeowners and auto markets have had actuarial scoring tools for decades — but the current wave of AI-first underwriting tools is primarily targeting commercial.
The vendors reviewed here are Gradient AI, Planck, Cytora, Sixfold, Federato, and Hyperexponential. All are specialist tools — none of them is a general-purpose software platform that also does underwriting. All are priced on a quote basis; pricing in this category is not publicly listed.
For context on how agentic AI is beginning to intersect with underwriting workflow automation, see our AI underwriting trends analysis.
Gradient AI is the most widely discussed AI underwriting vendor among commercial P&C carriers in the United States. Its core product is a loss prediction model that ingests the carrier's historical policy and loss data, external business data, and other structured inputs to produce a risk score for new submissions.
The genuine strength is the loss prediction accuracy for carriers with sufficient training data in concentrated segments — workers' compensation, commercial auto, and small commercial BOP are the lines most commonly cited by Gradient AI clients. The model is trained on the carrier's own data, which means it reflects that carrier's specific book characteristics rather than an industry average.
The limitation is the data requirement. Carriers with fewer than 50,000 policies in force in a given segment, or with loss data quality issues, will struggle to generate reliable model output. Gradient AI offers data augmentation services to address thin books, but the effectiveness varies. Before committing, conduct a data readiness assessment — Gradient AI's implementation team will typically do this as part of the sales process.
See our Gradient AI vs. Planck comparison for a direct feature contrast on data requirements and deployment complexity.
Gradient AI is also covered in our AI underwriting adoption report, which includes carrier deployment case data.
Planck takes a different approach than Gradient AI. Rather than building a loss prediction model on the carrier's historical data, Planck focuses on external data enrichment — gathering publicly available and proprietary business intelligence data about a submitted account and presenting it to the underwriter in a structured format.
For a restaurant submission, for example, Planck pulls liquor license status, health inspection records, online reviews, revenue signals, hiring data, and dozens of other indicators that speak to operational risk. The underwriter sees this context alongside the application, reducing the manual research time and surfacing signals that applications routinely omit.
The value of this approach is that it does not require the carrier to have deep historical loss data. Planck's enrichment is useful from day one, even for a relatively new MGA or a carrier entering a new segment. The limitation is that it does not predict losses — it enriches information. The underwriter still makes the pricing and accept/decline decision.
For small commercial BOP, workers' compensation, and light manufacturing risks, Planck's data sources are particularly well-developed. For specialty or complex risks, the publicly available data signals are thinner.
Compare Planck vs. Sixfold for a side-by-side of data enrichment versus generative AI summarization approaches.
Cytora approaches the underwriting problem from the workflow layer rather than the prediction layer. Its primary function is digitizing incoming submissions — extracting structured data from unstructured documents, emails, and broker portals — and routing them to the appropriate underwriting workflow based on risk characteristics.
The "risk digitization" framing is important. A significant portion of the inefficiency in commercial underwriting is not the pricing decision itself but the upstream work: manually re-entering data from broker submissions, chasing missing information, routing accounts to the right team. Cytora automates this intake and routing layer, which reduces the administrative burden before underwriting judgment is even applied.
This makes Cytora complementary to, rather than competitive with, tools like Gradient AI or Planck. A carrier might use Cytora to digitize and route submissions, and Gradient AI to score the routed accounts. The integration question becomes: does the routing system feed the scoring system cleanly?
Cytora's routing rules can incorporate appetite criteria, so submissions outside a defined risk profile can be declined or forwarded to a wholesale channel without reaching an underwriter's desk — a meaningful efficiency gain in high-volume small commercial. See our Cytora vs. Federato comparison for a detailed look at where the two tools diverge on portfolio management capability.
The concept of straight-through processing is relevant here — Cytora enables more of it for straightforward small commercial risks.
Sixfold is the tool in this category that most clearly reflects the current generation of large language model applications in insurance. Its primary function is reading submission documents — applications, loss runs, inspection reports, supplemental questionnaires — and producing a structured underwriting summary that the underwriter can act on.
The practical implication: an underwriter who would have spent 45 minutes reading a submission package can now spend 15 minutes reviewing a Sixfold summary and asking follow-up questions. The value is time savings per submission, which compounds when you are processing hundreds of accounts per month.
What Sixfold does not do is generate a risk score or make a pricing recommendation. It is a documentation and summarization tool, not a predictive model. This distinction matters when carriers are evaluating it — the business case is throughput and consistency, not loss ratio improvement from model-driven pricing.
The explainability profile is actually better than most predictive tools: if an underwriter questions a summary, the source documents are right there. The summary is auditable in a way that a black-box risk score is not. For regulatory purposes, this is an advantage.
Federato is unique in this category for focusing explicitly on the portfolio management layer. Its RiskOps platform does not primarily score individual submissions — it gives underwriting teams visibility into how individual decisions aggregate to portfolio-level outcomes.
The core problem Federato addresses: an underwriter approving a coastal property risk in isolation is making a sensible decision. An underwriting team that has already written 200 coastal properties in the same ZIP code concentration is making a portfolio risk that nobody has explicitly approved. Federato surfaces the concentration data in real time, so the underwriter pricing submission 201 has the portfolio context when making the decision.
This is a genuinely different value proposition than the other tools in this category, and for carriers who have experienced adverse portfolio concentration events, it is a compelling one. For carriers whose primary problem is pricing inconsistency on individual risks rather than portfolio management, Federato may be the wrong starting point.
Federato's RiskOps platform also supports agentic underwriting workflows, where routing and appetite rules are applied programmatically across the submission queue.
Hyperexponential is distinct from the other tools in this review in that its primary user is the pricing actuary, not the underwriter. The hx Renew platform is a pricing model environment that allows actuarial teams to build, deploy, and iterate on technical pricing models faster than traditional actuarial tools permit.
For standard commercial lines, this may be more infrastructure than is needed. For specialty and complex lines — marine, energy, financial lines, large property — where pricing is genuinely bespoke and model iteration is a competitive advantage, hx Renew addresses a real constraint. The actuarial modeling environment in most insurers is built on tools that were not designed for rapid iteration: spreadsheets, legacy actuarial software, custom code that only one person understands.
Hyperexponential replaces this with a collaborative pricing platform where actuaries build models that are version-controlled, testable, and deployable to underwriters in a usable format. The underwriter-facing interface shows the model output in context, with the ability to apply underwriting judgment as an overlay.
Pricing is quote-based and reflects the complexity of the deployment. This is not a tool for carriers writing standard BOP; it is for carriers with dedicated actuarial teams who have identified pricing model agility as a constraint.
When evaluating AI underwriting tools, four criteria cut through the feature-list noise.
Data requirements and readiness. Every predictive tool in this category requires data. Before evaluating vendors, conduct an honest assessment of what structured data you have, how clean it is, and how far back it goes. Vendors who do not ask hard questions about your data readiness in the first meeting are selling first and solving second.
Model explainability and regulatory defensibility. In most states, you need to be able to explain adverse underwriting decisions. Ask each vendor how their model output is documented, what the audit trail looks like, and whether the model has been reviewed for disparate impact. The explainable AI requirements are not going away. See also our audit trail glossary entry for context on documentation requirements.
Integration complexity. AI underwriting tools do not operate in isolation. They need to receive submission data from your policy administration system or submission portal, and they need to return output in a form that underwriters can use in their existing workflow. Ask for a realistic integration timeline — not the best-case scenario, but the median customer experience. Complex integrations with legacy policy systems routinely take longer than initial estimates.
Total cost including implementation. Licensing fees are only part of the cost. Data preparation, model training, integration development, and ongoing model monitoring are all real costs that belong in the business case. See our total cost of ownership glossary entry for a framework.
All tools in this category are priced on a quote basis. No vendor in this review publishes pricing on their website. Budget ranges vary widely based on the number of submissions processed, lines of business covered, and whether the engagement includes model training, integration services, or ongoing actuarial support.
Expect multi-year contracts and implementation fees separate from licensing in most cases. Pilot agreements with defined success criteria are available from several vendors and are worth negotiating before a full commitment.
InsurAItools is editorially independent. We do not accept payment for placement or rankings. Our evaluation methodology is described at /methodology.
Editorial verdict: For commercial P&C carriers with sufficient historical data and a specific loss ratio problem they can articulate, Gradient AI is the most proven choice. For MGAs or carriers entering new segments who need data enrichment rather than loss prediction, Planck is the more realistic starting point. Sixfold addresses a distinct but real bottleneck — submission review throughput — and is worth evaluating separately from loss prediction tools. Federato and Hyperexponential serve genuinely specialized needs (portfolio management and actuarial pricing, respectively) that are not substitutes for the others. Do not buy any tool in this category without conducting a data readiness assessment first.
It depends on the tool and how "small" is defined. Planck is designed to serve MGAs and small commercial carriers — the data enrichment model does not require you to have a large historical loss dataset. Gradient AI and Hyperexponential, by contrast, require substantial historical data and are priced accordingly. Most vendors in this category offer quote-based pricing, and some will structure pilot agreements. If you have fewer than 50,000 policies in force and limited internal data science capacity, many of these tools will be difficult to justify without a pilot.
At minimum, a meaningful AI underwriting model needs historical policy data (exposure, coverage, premium), loss data matched to those policies at the individual risk level, and enough volume to support statistical significance in the lines you are targeting. Five years of loss history with at least 5,000 to 10,000 claims in a given segment is a typical starting threshold cited by vendors. External data sources — business intelligence feeds, weather data, satellite imagery — can supplement internal data, but they do not substitute for it in loss prediction models.
Regulatory scrutiny of AI underwriting is increasing in most states. The core concerns are disparate impact (whether AI models produce outcomes that discriminate against protected classes), explainability (whether you can explain to a regulator or applicant why a decision was made), and the use of non-traditional data sources. Most vendors in this category have invested in explainability features, but the regulatory environment is not uniform — a model permissible in Texas may face challenges in California. Engage your compliance team before deploying any AI underwriting tool in production.
Marcus Reed is a senior insurance technology analyst with 12 years evaluating agency management and insurtech platforms. He previously served as operations director at a mid-size independent agency.