A mid-sized UAE-based, Abu Dhabi-licensed insurer writing approximately AED 400 million of gross written premium annually was confronted with an issue that many commercial insurers are only too aware of: the pace of underwriting was directly limiting growth. Brokers were expecting quotations immediately. The company was getting over 500 fresh submissions and renewals every month in the lines of property, marine cargo, motor fleet, group medical, workmen’s compensation, and liability lines. But underwriting relied on manual scrubbing of disorganized broker submissions, manual lookups on rating sheets stored in Excel and PDF files, and underwriter judgment that could differ significantly from person to person.
The result was slow turnaround, pricing inconsistencies, and too much reliance on knowing specific individuals. It would take two to three days to quote on a standard risk. Five to seven days for complex risks. Junior underwriters, at times, appear to price aggressively because they want to win business and damage the loss ratio. Others were too safe in their pricing and lost business opportunities. In broker submissions, missing information was frequently identified late, after a day or two of manual review. Reinsurance referrals were also slower than they had to be.
At aTeam Soft Solutions, we have developed an AI agent that reads applications, extracts structured risk data, leverages external intelligence, computes a technical premium baseline, prepares quotations, and forwards complex risks for human sign-off. The result was an underwriting process that was much faster, more consistent, and commercially more effective without the removal of underwriter judgment from the final decision.
As we observed the underwriting process, the initial thing that caught our attention was that the team was indeed engaged in real underwriting work; a majority of their time was being taken up by setup work as opposed to exercising risk judgment.
Each new commercial insurance application came in a slightly different format. Some appeared via broker email, complete with clean proposal forms, loss runs, and supporting schedules attached neatly. And some came in partial documents, scanned reports, handwritten notes, or attachments divided among multiple emails. One property risk submission might consist of a proposal form, survey report, building schedule, loss history, and photographs. A group medical case might have census information, employee demographics, claims utilization reports, and prior policy terms. A marine cargo enquiry may contain details of trade routes, commodity information, annual turnover, warehouse transit points, and clauses as requested.
An underwriter was required to normalize the submission before they could start the pricing process.
That involved going through the documents, determining the type of risk, extracting the relevant fields, verifying what was missing, figuring out which aspects of the submission were important for this particular line of business, and then deciding whether the data were sufficient to go on. This initial phase alone could take hours, particularly if the broker’s submission was incomplete or contradictory.
Then it was the turn of the underwriter to assess the true risk.
In the case of property risk, they analyzed building type, occupancy, fire protection, exposure concentration, location features, and history of claims. For the motor fleet, they evaluated vehicle mix, driver profile, patterns of use, claims history, and areas of operation. For group medical, they studied employee demographics and sector as well as utilization risk and prior medical side loss experience. For workmen’s compensation and liability, they reviewed industry risks, payroll, number of employees, site exposures, and legal climate.
There was plenty of work in that already, but the underwriter also had to reference rating manuals and internal pricing logic that existed between Excel files, PDF guidelines, and departmental knowledge. A few elements were well documented. The rest were mainly just through practice. Certain assumptions in the rating had not been updated for years. The terms of the reinsurance treaty added yet another layer. A risk may be acceptable at one premium level but not at another when treaty costs, retention limits, or referral thresholds are taken into account.
Since the company was dealing with various commercial lines, it was not as if a single pricing process was being applied over and over again. It was like a suite of related, but separate, pricing processes.
And this is the place where we started to get some inconsistency.
Senior underwriters had an implicit mental model about how to weigh off loss history against account quality, competitor pressure in the market, constraints imposed by the treaty or any other factor, and the value of the broker relationship. Junior underwriters didn’t have nearly as much of that context. Depending on who did the work, the same risk could result in a price difference of 15-20 percent. Much of that variation was reflected in good judgment. And some of it was reflected in inconsistency. Inconsistency in underwriting tends to manifest itself later in the combined ratio and does not immediately show up.
The problem of speed was equally serious.
Brokers are incredibly sensitive to turnaround times. In the UAE market, if an insurer takes more than two to three days to quote, it is likely the broker has already received multiple competing offers. So, the speed of underwriting is not simply a measure of internal efficiency. It has a direct impact on the conversion of new business. The client knew this and experienced it every day. Average risks took two to three days. Complex risks, particularly those requiring reinsurance assessment, may take five to seven days. By the time the quote was prepared, the market discussion had often moved on.
The team also spent its time on information requests. Missing information was often identified late in the process because someone had to read the submission manually before they realized the last five years of loss history were incomplete, a fire protection report was out-of-date, or a fleet schedule was lacking key information. That took brokers lost time waiting not only for the underwriter but also for the underwriter to tell them what had been missing from the beginning.
The impact on business was evident.
Loss ratio was running at 68%, a little above the target of 65%. That spread was not catastrophic, but it was a meaningful one. Price inconsistency and uneven risk selection were some of the reasons. The company was also losing good business because it could not respond with quotes as fast as the market wanted. And as the underwriting volume increased, the firm had a choice that many insurers are presented with: add more underwriters or revolutionize the underwriting process itself.
That is when aTeam Soft Solutions came into play. We were never asked to replace underwriters. We were asked to build an insurance pricing AI agent to perform the heavy preparation work, develop a technically more consistent baseline, and let underwriters use their time on judgment, exceptions, strategy, and broker negotiation.
The insurer already had an underwriting system, but it functioned more as a policy record-keeping tool than an underwriting engine. It could store policy information after decisions were taken. It wasn’t too helpful in rendering decisions.
The rating manuals were another constraint. They existed but were scattered across spreadsheets and PDFs. That still meant underwriters had to read, manually cross-reference, and apply them. In commercial insurance, that’s manageable at low volume. When a business wants to quote faster and have more pricing consistency between underwriters, that’s when that becomes a bottleneck.
The company had also considered a rules-based rating engine. The trouble was that fixed rules can only take you so far in business underwriting. Every real risk is an exception to the rule. A plant with poor loss experience but good fire protection is not the same as one with an equal loss ratio experience but poor maintenance standards. A commercially valuable broker may justify a closer commercial accommodation than a one-off transactional account. A hospitality risk might be rated differently if the carrier wanted to grow that class. Static rules do not account for such nuances very well.
That’s why this wasn’t a simple automation issue. It needed to be an AI insurance underwriting problem that would involve risk interpretation, understanding documents, extracting structured data, scoring logic, pricing discipline, and allowing space for human commercial judgment. The right answer was not “take the underwriter out of the loop.” It was to develop an AI agent that could prepare, analyze, and recommend at speeds no human processes could allow.
We developed the solution in stages because the underwriting teams must trust the baseline before it is allowed to affect quote decisions. We started with intake quality and data structure, then moved into risk scoring and pricing logic, then quotation automation and, ultimately, portfolio-level intelligence.
Phase one concentrated on the chaotic front end of underwriting, which was losing too much time prior to actual underwriting having even begun.
When a submission was received through broker email, direct portal, or any other intake channel, the AI agent consumed all the materials attached and processed them as one risk package. It pulled out the relevant information based on line of business: company profile, turnover, asset schedules, building traits, fleet makeup, employee numbers, claims history, survey observations, and other structured underwriting data.
It all sounds straightforward until you actually see the documents mix. Proposal forms, PDFs, scans, spreadsheets, survey reports, financial statements, narratives written in emails, they are all part of the case. The AI agent needed to merge those fragments into a single clean view for underwriting. That’s what made the underwriter’s job infinitely easier.
We have also added external context to the application where it makes sense: credit signals, records of public companies, geographic and hazard context, and industry benchmarks relevant by class of business. That did not replace underwriting judgment. It gave it a more solid factual base.
One of the most important functionalities during this stage was the on missing information. Instead of waiting a day or two for an underwriter to discover that a key item was missing, the system flagged missing pieces on the spot and issued a specific request to the broker. Two aspects helped turn it around: underwriters weren’t wasting time reviewing incomplete files, and brokers were learning what was missing at the beginning of the process, rather than the end.
When a case made its way to an underwriter, it was no longer just a collection of paperwork. It was a risk summary, formatted and structured.
When the intake is steady, we step into the core underwriting logic.
The AI agent analyzed loss history patterns, compared the risk to the insurer’s historical portfolio, found positive and negative risk attributes, and produced a structured risk score. But significantly, the output was more than just a number. It covers a rationale.
For example, in a property case, the system might indicate a clean loss history, sprinkler coverage, low hazard occupancy as favorable factors, with an older profile electrical installation or a geographic concentration risk as factors that increase the score. On a fleet case, the emphasis might be on vehicle type mix, historical frequency, and route exposure. For group medical, it could include workforce age mix, industry, utilization history, and previous claims cost trends.
Starting from this, the AI agent used the appropriate rating methodology, constructed a technical premium baseline, tested it against treaty constraints and internal minimums, and output a recommended pricing range: a lowest acceptable level, a target rate, and a competitive rate band.
This is where a person’s judgment was still needed.
At aTeam Soft Solutions, we understood from the beginning that underwriting is not just math. Market sense, broker relationship value, strategic growth sectors, and potential at the account level for cross-selling—all these things matter. So we develop what we call a judgment overlay. It gives the data-driven technical answer. The underwriter may then make strategic adjustments to that view, but those changes are explicit and documented. That means the company maintains pricing discipline without acting as if every quote should be handled as a machine exercise.
This stage drastically reduced the quote preparation, as the underwriter was no longer beginning from scratch. They began with a structured, technical position and aimed at commercial decision-making.
Once the technical pricing level was functioning properly, we transitioned into quotation production.
In the case of the usual risks that were within the normal parameters and within the pre-agreed authority, the AI agent would produce the quotation automatically in the insurer’s branded format, along with its core terms and conditions, the premium, and a summary of the coverage. If needed, it also added the required disclosures and references of the policy wordings relevant to the UAE regulations expectations.
The turnaround was dramatically altered. Routine risks that in the past took two or three days could now be taken through to the broker-ready quotation stage in a matter of hours, provided the submission was complete.
In more complicated or special risks, there was no attempt to over-automate by the AI agent. Rather, it created a solid risk presentation and pitched it to the underwriting manager or reinsurance team with the analysis already assembled. It meant that the senior reviewer was not reading from nothing. They were considering a prepared case.
We also automated a portion of the broker interaction loop. If the broker did not reply within the allotted time, the AI agent sent a follow-up. If the broker pushed back on premium, the system could determine if a reduction was allowed by risk and treaty parameters, and if it could, it would suggest a response route to the underwriter. This was particularly helpful because a great deal of quotation slippage in insurance is not derived from the first offer but from a slow handling of broker negotiation after the initial offer is sent.
By this point, the insurer operated what was effectively a standard commercial flow automated underwriting system, though individuals still made the final underwriting decision when necessary.
The final major stage was really expanding the system from just individual quotes to portfolio intelligence.
The AI agent began tracking the insurer’s book for exposure concentration, segment profitability trends, and evolving loss signals. It could sense where excess capacity was accumulating in a particular geography, one industry, or one treaty-sensitive class. This provided underwriting leadership with far better insights as to whether their online quote decisions on a daily basis were adding up to the portfolio they really wanted.
We also applied the same principles for renewals as well, which turned out to be one of the most commercially valuable parts of the platform. The insurer had over 1,500 annual renewals, and the AI agent was able to use claims experience, likelihood of retention, market conditions, and value of the account to recommend renewal pricing that balanced retention and profitability. The renewal retention has improved since the company wasn’t so slow or inconsistent in its renewal reply.
Recommendations for preventing loss were another helpful by-product. If the system detected worsening claims patterns or developing stress in an industry, it could bring forward clients requiring attention before renewal or loss escalation that can become severe. It was that which gave the platform not simply a speedier quote engine, but a far more powerful underwriting intelligence layer.
The backend was developed in Python using FastAPI, with Celery and Redis managing asynchronous processing of submissions, scoring jobs, document extraction, quote generation, and portfolio analytics. Structured underwriting data, pricing outputs, rule states, and quote history were stored in PostgreSQL.
Azure AI Document Intelligence handled most of the OCR and field extraction on the application form, and Claude was responsible for risk narrative generation, document reading, and the explanation layer that made the recommendations understandable to the underwriters. Much of the risk scoring was powered by XGBoost models, trained on five years of policy and claims data, where there was enough volume.
For low-data lines, we purposely did not force pure ML. Instead, we applied a hybrid approach that included industry benchmarks, rules, and severity analogs from related classes. That mattered because the commercial insurance portfolios are typically uneven by class. Certain lines enjoy rich historical data. Other lines do not. A good AI agent needs to understand when to rely on ML and when to retreat to structured actuarial logic.
We integrated the platform with the client’s legacy underwriting system via database connectors, as replacing the entire core system would have been unwarranted and risky. The React dashboard became the underwriter’s working surface, showing clean risk summaries, extracted submission information, pricing baselines, judgment overlay controls, referral triggers, and portfolio context.
This architecture allowed the system to act as a true agentic AI insurance UAE workflow instead of a scoring widget.
One of the largest challenges was the disconnect between technical pricing and business reality.
Underwriters with experience are not pure price calculators from formulas. They consider broker relationship momentum and strategic sectors the company wants to grow into, competitor behavior, and even account profitability, not just one policy. If we had imposed a stiff “AI sets the premium” approach, it would have been thrown out. That’s why the judgment override was so critical. It allowed underwriters to make strategic adjustments while maintaining a uniform technical foundation.
Low-data insurance lines present another difficulty. There was insufficient portfolio history for certain business classes to allow a strong ML model to be built. We fix that by going with hybrid logic instead of pretending all classes were equally data-rich. It made the system more believable because it did not overstate certainty where the data were insufficient.
Reinsurance treaty rules were also more complex than many people anticipate. The insurer had quota share and excess-of-loss arrangements with limits, exclusions, minimum rates, and referral thresholds that directly influence pricing. We encoded more than 50 treaty rules so the AI agent can prevent underwriters from pricing below treaty economics or binding business that requires the reinsurance review.
The last challenge came post-go-live. The processing speed got so high that underwriters started approving the suggestions too fast at the beginning. This is a classic, but important, issue in an enterprise. AI: The output getting faster can accidentally weaken human oversight if the reviewing discipline does not evolve. We responded by introducing a required review checkpoint for the underwriter to confirm a small number of key risk factors prior to approval. That’s just a few minutes more, but it keeps the quality of underwriting.
The speed improvement was instant and worthwhile commercially. The standard quotation turnaround time reduced from 2-3 days to less than 2 hours. That instantly changed the brokers’ perception because the insurer was now competitive on response time in a market where speed often decides who even gets to remain in the conversation.
The capacity to underwrite increased enormously. The same group of eight underwriters was now able to process approximately five times the number of submissions, as the burden of administration and preparation had been so greatly reduced. That enabled the insurer to expand without increasing underwriting headcount at the same pace.
Also, pricing consistency became better. Divergence between writers on the same risks went from 15-20% down to less than 5%. The reason this is important is that consistency in pricing is one of the strongest signs of a more disciplined underwriting operation. It also surged through to profitability: the loss ratio improved from 68% to 63% at the first renewal cycle, propelling the firm beyond its initial goal.
Renewal retention increased from 82% to 90%, in part because renewals were processed more quickly and with more responsiveness. Gross written premium increased 22% without the need to hire additional underwriters, and broker satisfaction improved as quote response speed was approximately 80% better than previously.
Handling of information requests was also improved in a way that provides the immediate delight of brokers. Requests for missing information are now sent out nearly immediately, rather than after a day or two of manual review. That led to a much smoother experience for the broker and less wasted time back and forth.
Reinsurance referral handling was also enhanced as well. Files that once sat for five to seven days were now completed, in many cases, within two days for reinsurance review, as the AI agent had already put together the presentation and supporting logic.
For the client, one of the biggest secondary benefits was that it helped with regulatory and management reporting. The system produced many of the needed compliance and portfolio monitoring reports automatically, reducing the workload of the compliance department and providing management with much better visibility into underwriting quality at the book level.
In the end, the project demonstrated that a powerful insurance pricing AI agent can enhance growth, stability, profitability, and pace of operation all at once—but solely if it’s built to empower underwriters instead of bypassing them.
We developed the solution in Python using FastAPI for backend orchestration, Celery and Redis for asynchronous processing, Azure AI Document Intelligence for form OCR and extraction, Claude for document understanding and underwriting narrative generation, XGBoost for risk scoring, PostgreSQL for underwriting and quote data, React.js for the underwriting dashboard, legacy-system integration via database connectors, and AWS EC2, RDS, S3, and SageMaker for hosting, storage, model training, and processing support.
The main takeaway is that speed and control need to be co-designed together.
During the initial stage after rollout, faster quote generation introduced an unexpected hazard: people started trusting the output so quickly that they occasionally gave it less careful consideration. That is not a model failure. It is a design workflow problem. After introducing the mandatory review confirmation step layer, the quality of human review stayed robust, and the majority of the speed benefit was not lost.
And what we also learned is that in commercial underwriting, the winning model is not “AI replaces underwriter”. The model that wins is that the AI produces a technically sound, consistent baseline, and the underwriter brings commercial judgment. That framing increased trust, and portfolio discipline improved at the same time.
And finally, we found that underwriters were quicker to adopt AI when the system told them why it arrived at a pricing view. The technical premium alone is not enough to convince. Underwriters want to know what risk factors contributed to the recommendation and how it compares to the insurer’s portfolio. This explanation layer became a key to adoption, and today one of our core values at aTeam Soft Solutions is for any serious underwriting automation development India engagement in regulated financial services.
Commercial insurers throughout the UAE and the wider Middle East will commonly know they have a problem with speed, a problem with pricing consistency, or a problem with renewal efficiency. What they rarely underestimate is how connected those issues really are.
At aTeam Soft Solutions, we develop platforms in which an AI agent can normalize submissions, score risk, compute technical premiums, produce quotations, and provide portfolio intelligence — all the while allowing the underwriter to retain control of the ultimate commercial decision. That’s how underwriting gets faster without getting weaker, and more consistent without getting rigid.
