Using AI Assistants

Once your assistant is connected, you can ask it billing questions in plain English. This page covers what works well, how the assistant sees your data, and how to keep your connections tidy.

What to Ask

With reporting KPI access, the assistant can answer aggregate questions:

“Show invoice totals by month for this year.”
“How is our overdue debt spread by age?”
“How many Direct Debit collections failed last month, and for how much?”
“How does this quarter’s customer growth compare with last quarter?”
“What share of payments came in by card rather than Direct Debit?”

With individual record (read) access, it can also work with specific records:

“Find the customer account for Smith’s Bakery in Leeds.”
“List their unpaid invoices, largest first.”
“Which customers added in the last 12 months have overdue invoices?”
“Show me any open tickets for account 10234.”
“Give me a summary of this account.”
“What is on invoice 12345, and which charges are the biggest?”
“What did that payment settle?”
“Why did their card payment fail last month?”
“What did each of these numbers bill last quarter?”
“What rate are we charging them for mobile calls?”
“If I change the national rate on that tariff, how many other tariffs does it reprice?”

Where the CRM is in use, it can also answer sales and credit control questions:

“Which deals are closing this month, and what are they worth?”
“How many cases are breaching their SLA targets?”
“How much overdue debt have we recovered after contact this quarter?”
“Which proposals are still with the customer and past their reminder date?”

With write access, it can act, using the platform’s own workflows:

“Email that invoice to the customer’s accounts contact.”
“Add a note to the account summarising this conversation.”
“Create a TODO for me to review their tariff next month.”
“Suspend that number.”

Good assistants will confirm significant actions such as drops, write-offs and refunds with you before performing them, and will ask which effective dates you want rather than assuming. The platform steers them to do exactly that, but it cannot force a badly behaved client, so word your requests precisely when acting on records.

With call data access, it can answer traffic and cost questions:

“How many calls did we carry last month, and what did they cost us wholesale?”
“Break international minutes down by month for this year.”
“Which dealers sent the most traffic last quarter?”
“Show the calls on that invoice, longest first.” (needs full per-call detail)

Aggregate summaries cover up to 366 days. They can group calls by period, direction, call type, rating period, number type, dealer, carrier or CDR file type. You can also ask for one carrier’s traffic.

Summaries for a specific customer, number, service, feature, invoice or transaction also need individual record access. Both call grants are read-only, so an assistant can report on calls but never change or delete them.

Grouping by site or by number needs individual record access as well, because both identify the customer rather than describing traffic: a site is free text typed against the number and routinely names the account. Dealer is not treated this way, being a controlled code shared across many customers.

Cost figures cover every leg of a call, including the termination charges on the B leg, so the call costs an assistant reports for an invoice reconcile with the invoiced call charge. Call counts, minutes and data usage cover A legs only, which is the count you want when asking how much traffic was carried.

Working with billing runs

With Billing Operations read access, an assistant can review billing runs and CDR files. Ask it to:

“Which checks are outstanding before we request this month’s bills?”
“Which carriers have missing or short CDR files?”
“Show the Billing Preview and customers whose projected invoice differs most from their previous bill.”
“List the missing retail tariffs we need to resolve before billing.”
“How many invoices still need emailing, printing or collecting?”

The assistant reads the billing run’s checks and progress, including confirmations, pending files and billing errors. The preview compares projected charges with previous invoices and keeps totals separate by currency.

Billing-error results follow your View Errors permission and Billing Error Reports settings. Sections naming numbers, customers or tariffs also require access to those records. The platform’s MCP settings user must permit the same access.

With Billing Operations write access, an assistant can perform the preparation tasks your permissions and the run’s status allow:

Confirm checks you have reviewed, including payments, CDRs, billing errors and reports.
Ignore missing or pending CDR files when you explicitly choose to continue without them.
Request pre-billing or post-billing reports, request bills, or return a requested run to re-check.
Check CDR email and FTP sources, process pending files, or re-recognise them.
Reset, hold or re-recognise an eligible individual CDR file.

For example: “I’ve reviewed the pre-billing report. Confirm that check for this run.” Review the evidence before asking an assistant to confirm a check or ignore missing files.

Run Bills, Approve and Ready remain staff actions in the platform. Requesting bills through an assistant follows the same workflow as requesting them from the billing run page.

Comparing carrier traffic

Ask “Compare each carrier’s monthly traffic with its usual levels and highlight missing files or unusual costs.”

Carrier traffic history follows the Pre-Billing Report. It shows monthly files, call legs, minutes, costs and profit against rolling averages, with the report’s warnings. It also identifies expected carriers with no calls in the period.

This requires Billing Operations read and Call Data aggregate access. For a simpler comparison, ask for call summaries grouped by month, carrier and CDR file type.

Working from a list

An assistant can look up many values at once rather than one request per row, which makes list-shaped questions practical:

“Here are 80 numbers from a porting notice. Which customers do they belong to?”
“Check these 50 account numbers and tell me which have unpaid invoices.”

Up to 100 values can be checked in one go against customers, numbers, services, features, invoices, payments, payment instruments, contacts, tickets and notes. Results stay grouped against the value you supplied, and where a value matches more than one record the assistant is given the alternatives to resolve rather than a silent guess.

Number lookups behave like SmartSearch: formatted and protected numbers are understood, displayed and CDR variants are both searched, and exact matches come first.

Summaries and rollups

An assistant can ask for a whole picture in one operation rather than assembling it from separate lookups.

A customer summary covers the account balance, the overdue amount and total exposure, the billing numbers, services and features, recent issued and unsent invoices, recent payments, and open TODOs. An invoice summary covers the dates and totals with the charges grouped by where they came from and the largest charges called out. Each section still obeys your own permissions, so a summary shows the parts of the picture you could see yourself.

With call summary access, an invoice summary also breaks the calls down by call type, and by number or feature where your permissions extend that far, without returning individual call identifiers or called parties. With discount plan access it shows the allowances that applied in the billing period, separating what was granted, used on this invoice, used to date and left over, from the figures stored at the time rather than today’s defaults. With full per-call detail, the assistant can rank an invoice’s calls by charge, base price, price difference or date, and explain how any one of them was priced.

Grouped counts and totals answer “how many” and “how much” across a filtered list without paging through it: numbers grouped by customer, type, status, dealer or the month they were sold or dropped, or transactions grouped into their recurring charges, one-off charges and fees, call charges, refunds and combined value. Each group returns its count, the number of distinct customers in it, and a few sample records to check the grouping is what you meant. These are record-level tools, so they need individual record access, and a per-number call rollup needs number access with them.

Money in a grouped result is split by currency, so a grouping that spans currencies returns a group per currency rather than one meaningless total.

Unlike the KPI endpoint, which is aggregate-only, an assistant can scope these figures to a single customer - it acts as you, so your own customer access applies. That needs record access to the customer as well as reporting access, and the Personal Customers grant where the customer is a personal one.

Reporting questions can also turn on a threshold or a range, rather than only on the categories a KPI already groups by - invoices over an amount, collections retried three times or more, customers on a particular billing cycle. The assistant filters on the same fields the KPI endpoint permits for that category, so these stay aggregate questions answered by a single query rather than by listing and counting records.

The rates an assistant quotes

Asked what a customer pays for a call type, an assistant reads the effective rate sheet: the same view the tariff, number and customer screens show, with based-upon inheritance, the tariff’s base defaults, dated rate versions and fallback rates already merged in, and the same human-readable Rate Summary. The alternative - the rate lines defined directly on a tariff - is a delta stack, and for an inherited or fallback-priced call type it holds nothing at all, so it is not where a pricing question is answered.

It can read the sheet three ways, matching the three grids an operator uses: a tariff’s whole sheet, the rates one number gets from its resolved tariffs at its price-lock date, or the tariffs in use across a customer’s numbers together with its default tariffs. The number and customer views start from the headline call types as those grids do, and fall back to the whole sheet where a platform marks no headline types, saying which it applied.

Rates are quoted in minor units - pence for sterling - and carry no VAT, discount plan or allowance. Reading the sheet at all needs record read access; the full set of rate columns needs the same tariff rate display access the record tools require, and a connection without it gets the compact set.

Each rate says where it came from, so an assistant can tell an inherited rate from one defined on the tariff being read, and is told how many other tariffs are based upon each tariff involved. That is what lets it say “changing this reprices seven other tariffs, adding one here reprices none” before it writes. Nothing is blocked: repricing a shared tariff is a legitimate instruction, and the platform accepts it from that tariff’s own page too. A write over the connector onto a tariff others are based upon is logged with the tariff and the number of tariffs repriced.

Checking what a connection can do

Assistants can ask the platform who they are acting as and what they are permitted to do, so a well-behaved one will tell you it lacks an access category rather than guessing at why something failed. If an assistant reports a missing capability or a problem, it can also send that feedback to the platform maintainers directly.

How the Assistant Sees Your Data

A few things worth knowing when results look abbreviated:

Summaries first. Record lists return a compact headline view (account number, name, status, and similar); the assistant fetches full record detail only when needed.
A customer’s money position is read, not added up. The balance, overdue amount and exposure come from the customer record itself, as do the average and lifetime spend figures, lifetime payments and the monthly equivalent of the recurring charges, so they match the platform’s own screens rather than an assistant’s arithmetic over invoices. Each figure needs access to what it is derived from, and that requirement applies to the platform’s MCP settings user as well as to you: a connection whose MCP user is denied Invoices cannot read invoice-derived figures even where you hold Invoices yourself. See Financial figures on a customer.
Recurring charges come from feature records. A customer’s actual rental is the charge on their feature, alongside the interval it is charged over, so monthly, quarterly and annual amounts are not silently averaged together. Where a feature is priced by a fixed fee tariff rather than an explicit amount, the assistant sees the resolved charge rather than a blank. Fixed fee tariffs are the default price book, and transactions are the billed lines for a particular invoice or period.
Result caps. KPI queries return up to 200 rows by default (500 at most); record lists return up to 25 records by default (100 at most). Responses tell the assistant when results were truncated, and it can page through for more.
Personal customers may be out of scope, in totals as well as records. Unless the connection was granted Personal Customers, individual records are limited to customers classified as businesses, so an account you can see in the platform may not appear in an assistant’s results. KPI totals cover the same business-only population, so a figure from such a connection is not a platform-wide figure. The assistant is told which basis it is quoting; ask it if it does not say. The KPI endpoint always reports platform-wide, so the two legitimately differ.
Your permissions apply. The assistant sees only what your own account can see, including customer scoping, and only within what your platform’s MCP settings allow. If it reports that access is not enabled, your data controller needs to grant the relevant Access MCP flag, or the connection needs re-approving with that access ticked. See What you can grant.
Headline results follow list permissions; full detail needs view permission. A batch lookup can therefore tell you a record exists and give you its headline fields while still withholding the detail, if that is what your permissions say.
Call charges are explained from stored evidence. When an assistant explains how a call was priced, it reads the tariff, rate row and discount decisions recorded at the time of rating, rather than inferring them from the call’s current state. See Rating evidence.
Everything is logged. Actions taken through an assistant appear in the platform’s activity log against your user, flagged as MCP activity so they are distinguishable from your interactive and API activity.

Managing Your Connections

Open the AI Assistants (MCP) page from the Advanced menu. The Your Connections table lists each connection your account has approved, showing when it was authorised, when it was last used, the access it holds, and its status (Active, Expired, or Revoked).

To cut off an assistant, choose Revoke next to the connection. Revocation is immediate: the assistant can no longer access the platform with that connection, and reconnecting requires a fresh approval with your login and 2FA.

A connection that has not been renewed for sixty days expires by itself.

If You See a Connection You Do Not Recognise

You receive an email whenever a connection is approved or revoked with your account. If you receive an approval email you did not initiate, or spot an unfamiliar connection in the table:

Revoke the connection immediately.
Reset your password and 2FA, since someone may have your credentials.
Tell your data controller or support so they can review the activity log.

AI Assistants overview - what a connected assistant can do and how access is controlled
Initial setup - connecting Claude, ChatGPT, or another assistant
KPI Module - the reporting data behind the assistant’s aggregate answers
CRM machine access - the CRM record families, the sales and credit control access levels, and the /crm/mcp address