AI for Claims Handling in Telecom
A contractor opens a footway on a Tuesday morning, puts a bucket through a duct, and a residential street loses service until late evening. Elsewhere in the same operator, an enterprise customer claims credits against an availability commitment, a wholesale analyst contests an interconnect invoice from another carrier, and a retail adviser takes a cracked handset into a warranty process the operator does not own. All four are called claims. None of them sit on the same desk, run on the same clock, or need the same evidence.
Six different jobs filed under one word
An operator is almost never an insurer. Before any of this can be automated, the word has to be unpacked.
Outbound damage recovery puts the operator in the claimant's chair. Third parties damage buried and overhead plant constantly, and the operator pursues them for reinstatement and for what the loss of service cost.
Inbound service credit claims put the operator in the payer's chair. A customer on a contracted availability commitment says the service fell short and asks for money back against a schedule.
Wholesale and carrier disputes are fought with another operator over interconnect, roaming or capacity billing, under a dispute process with its own notice periods and formats.
Handset insurance and warranty are usually distributed rather than underwritten. The operator's job is intake and handoff, and its risk is the customer believing the operator decided.
Wayleave, access and property damage claims arise from the operator's own works on somebody else's land.
Subrogation runs the other way again, when another party's insurer pursues the operator for what it has already paid out.
Those six live in network operations, legal, wholesale, retail and property. They share a word and very little else, and nobody has ever seen them laid out together.
A contractor's bucket through a duct
Take the damage recovery book, where the money leaks most quietly. A recovery file needs the fault ticket that opened when service dropped, the network records showing what plant was where and who it served, the plant location enquiry and the locate records proving what was issued and when, the contractor's own drawings if you can get them, photographs taken at the dig, a reinstatement cost built up from labour, plant and materials rather than asserted as a round total, witness detail from whoever was standing there, and a recovery letter putting all of it in front of the third party.
Every one of those items already exists. They sit in six systems that were never joined, held by teams with no reason to talk to each other, and get pulled together by hand, slowly, usually well after the limitation clock has started moving.
Pulling a recovery file out of six systems
This is retrieval and reconciliation, which is work an agent suits. Given a fault ticket reference, it can pull the matching network records, the enquiry and locate history for that location and window, the photographs attached to the job and the cost build-up from the repair crew's record, then present them as one pack with every item sourced back to its system.
The missing list is the useful half
The pack matters less than the list of what is absent. No locate record retrieved. Photographs present but undated. Reinstatement cost supplied as a single figure with no breakdown. A handler who knows that early runs the claim differently from one who finds out when the other side's adjuster points it out.
From a complete pack it can draft the recovery correspondence in the operator's template, quoting only evidence it holds. A person signs it.
The claim against you starts with a schedule, not an outage
Turn the desk around. An enterprise customer claims credits because a circuit was down. The instinct is to check whether it was, and that is rarely the question that decides anything. What decides it is the contract schedule: how availability is defined, what is measured, over what period, whether the clock starts at detection or at the customer raising it, what the credit is, and what is carved out. The claim is tested against the document, not against the memory of a bad week.
The outage happened and the claim still fails
Most contested service credit claims turn on an exclusion rather than on uptime. Planned works properly notified. Force majeure. Customer-owned equipment. Power at the customer site. A third party denying access to a building so an engineer could not reach the fault. A fault the customer sat on for days where the contract requires prompt notification.
Each is answerable from records the operator already holds: the monitoring data, the fault ticket and its timestamps, the notification record, and the schedule. An agent can test the claim against all four and produce a reasoned position naming the exclusion it turns on, with the evidence for and against set out rather than summarised away.
Works that were notified to nobody in particular
The common failure is an outage logged internally as planned where the notification never properly reached the customer, or went to an address nobody reads, or arrived inside the notice window rather than before it. Defending the claim, you find this out late, after the position has been asserted in writing. Cross-reading the maintenance record against the notification log is dull work an agent will do every time, which is the point.
A dispute note from another operator, in their own format
Carrier disputes have their own etiquette: a dispute window, a required format, a reference convention, and a counterparty whose analyst is doing the same job from the other side. The substance is reconciliation of contested minutes, events or capacity against the usage record and the rate card in force on the dates in question.
An agent can line the disputed invoice up against your own usage and rating data, isolate the lines where the two genuinely disagree, separate those that differ only by a rounding or rating convention, and produce the submission in the counterparty's required shape. What to concede stays with the person who owns the relationship.
Cover you distribute and do not underwrite
Handset insurance and warranty sit awkwardly because the operator is the shopfront for somebody else's decision. The work is intake quality and clean handoff: capture the fault, the device, the proof of purchase and the circumstances once, properly, and pass a complete referral to the underwriter or manufacturer.
The real exposure is reputational: when the claim is declined, the customer remembers being in your store, talking to your adviser. An agent can make the referral complete first time and keep the customer told where the file sits, which removes most of the complaint volume. It must never imply a decision that is not yours to make.
Damage your own crews caused on somebody's land
Civil works on private land generate the mirror-image claim: a damaged driveway, a broken boundary wall, a landowner who says the wayleave never covered that route, a farmer with crop damage. They arrive as letters and phone calls rather than forms, handled by whoever answered.
The useful automation is recovery run backwards. Match the complaint to the works programme, the permit, the wayleave or access agreement and the crew's own job record, then set out what the agreement permitted beside what the record says happened. Where those two disagree, a person needs to go and look.
A letter from an insurer acting for somebody else
Subrogated claims are easy to misfile. They look like correspondence rather than claims, and the sender is an organisation you have no relationship with, pursuing a loss it has already paid its policyholder.
An agent reading inbound post can spot the subrogation shape, identify the underlying incident in your records, and open the file against the right event instead of letting the letter sit in an inbox. Half the damage comes from not recognising them in time.
One calendar for everything that has a date on it
Across all six books there are dates that bite: the limitation period on recovery, the notice window in a carrier dispute, the response deadline in a solicitor's letter, the period inside which a customer must claim a credit, the reply the regulator expects. They live in different systems, owned by different teams, and nobody holds the whole set.
One watch across all of them, counting down and escalating before a breach rather than reporting one afterwards, is the least glamorous item here and probably the highest value. Missing a limitation date converts a strong file into nothing.
What is open, against whom, and waiting on what
Ask most operators how many claims are live, their value, and which are waiting on them rather than on someone else, and the answer takes a week of asking around. One view across the six books, showing counterparty, stage, next date and the specific thing each file is blocked on, turns a set of private queues into something manageable. It also makes patterns visible: the same contractor striking plant repeatedly, or the same route generating credits every quarter.
Where AI fails here
Admitting liability and settling are human authority and should sit with named people, logged as their decision. No agent should be positioned to concede a point in writing.
An evidence pack is not a legal opinion. It arranges what you hold and marks what you do not. Whether a duty was owed and breached belongs to a lawyer.
Monitoring data can disagree with what a customer experienced, and that disagreement is a conversation, not an output. Telling a customer their bad fortnight did not happen because a graph says so is how a credit dispute becomes an escalation.
A disputed exclusion that turns on what a notification said will turn on judgement. Whether a particular email was proper notice under the schedule is an argument, and arguments need a person willing to own them.
Litigation and regulatory escalation decisions are not automatable, and should not be framed as recommendations.
Pursuing a small claim is sometimes the wrong commercial choice even when the file is strong, because the counterparty is a supplier you need next quarter. A system that scores files purely on strength will push you into fights you should avoid.
Where to start
Take the damage recovery book, and inside it the single most common cause of plant damage on your network. Automate three things for that one pattern: file assembly with sources named, the missing evidence list, and the limitation countdown. Measure how long a handler now takes before they can decide to pursue, and audit whether the missing list was right.
When that holds, add the service credit test against the contract schedule, because it is the same discipline pointed in the opposite direction. Adopt AI Partners works in that order deliberately, since recovery is where the unclaimed money sits and the evidence is already yours. If it would help to talk it through against your own books, hello@adoptaipartners.com reaches us.