The agent cannot read what it needs
Order state, asset detail and contract terms sit across systems the agent has no grounded access to, so it answers from generic knowledge. Confidently, and often wrongly, which is worse than not answering.
Salesforce · Communications Cloud
An agent that cannot act is a chatbot. The difference is your data model.
Agentforce for Communications brings prebuilt, telecom-specific AI agents to the Salesforce platform, connected to CRM, OSS and BSS data so they can read live customer, service and billing state and act on it. Whether they work depends almost entirely on what they can read and what they are allowed to do. Twopir Consulting does the unglamorous part: grounding, action design, guardrails and measurement. The model is not the hard bit. The data model is.
12+
500+
40+
250+
Trusted by 500+ organizations — including connectivity and communications providers preparing their Salesforce data for agentic AI with Twopir Consulting.
Where We Work
- Salesforce Partner
- AI Delivery
- Agentforce
- Data Grounding
- Action Design
- Guardrails & Escalation
- Containment Measurement
- Agent Observability
Six reasons a telecom AI pilot never leaves the pilot
Almost none of these are model problems. They are data, permission and measurement problems wearing an AI label.
The rules live in code, not in the model
Eligibility, compatibility and pricing logic buried in Apex is invisible to an agent. Rules that live in a well-modelled catalog are readable; the same rules in code are not, which quietly caps what any agent can do.
It can explain but not act
Without designed actions the agent summarises a problem and hands it to a human, which moves work rather than removing it. Deflection measured this way looks positive and changes nothing.
Nobody defined what it must not do
Guardrails are added after the first bad outcome. In a regulated, billing-connected environment that is an expensive way to discover where the limits should have been.
Success is measured by conversation count
Volume of interactions tells you the agent was used, not that anything improved. Without containment, resolution and escalation rates against a baseline, the pilot cannot be judged.
The agent is a channel of its own
Built outside the flows that people already use, it produces a fourth version of the truth — so a customer gets one answer from the agent, another from care, and a third from self-serve.
What Agentforce Is Here
Prebuilt agents, grounded in your own stack
Agentforce for Communications is Salesforce's set of prebuilt, telecom-specific AI agents. They run on the Agentforce platform and connect to CRM, OSS and BSS systems so they can read live customer, service and billing data and take action on it, rather than answering from a general model. Salesforce has announced five: Billing Resolution, which analyses bills across fragmented third-party data and helps resolve disputes within defined guardrails; Quoting, which generates complex telecom quotes from predefined business rules; Site Grouping, which organises multi-location deals and suggests configurations from natural-language prompts; Service Level Objective Insights, which compares real-time network usage against compliance criteria for commitments such as fibre uptime or repair speed; and Guided Selling, which helps field technicians identify upsell opportunities during onsite visits.
The thing worth understanding before buying any of it: a prebuilt agent is a capability, not a solution. Each of those agents is defined by what it can read and what it is permitted to do in your org. A Billing Resolution agent is only as good as its access to your billing and usage data and the guardrails that say which adjustments it may make unaided. A Quoting agent is only as good as the catalog model behind it — if your eligibility and pricing rules live in Apex rather than in the product model, the agent cannot see them, and it will either refuse to quote or quote something wrong.
That is why most of our work on this page is not AI work. It is grounding: making order, asset, contract, billing and catalog state readable and trustworthy; action design: defining what the agent may actually do and what happens at the boundary; guardrails: limits, approval thresholds and escalation paths decided before deployment rather than after an incident; and measurement: containment, resolution and escalation rates against a real baseline. Twopir Consulting is a Salesforce Partner and delivers AI work as part of its practice; Salesforce builds and licenses Agentforce itself.
Freshness — verified September 2026. Of the five agents above, Salesforce described Billing Resolution, Quoting and Site Grouping as available at the time of writing, with Service Level Objective Insights and Guided Selling expected later. Agentforce moves quickly and that split changes; confirm current availability and licensing against Salesforce's own product pages before making a commitment based on it. We re-verify this section on a fixed cycle rather than leaving it to drift.
Five questions that decide whether an agent will work
We ask these before proposing any agent use case. A no in the first two columns is not a reason to abandon the idea — it is the actual project.
Agent readiness, question by question
| Question | Why it decides the outcome | What we do when the answer is no |
|---|---|---|
| Can the agent read the real state? | An agent without grounded access to order, asset and contract state answers from generic knowledge — confidently and wrongly | Build the grounding: surface the records, fix the access, connect the missing system |
| Do the rules live in the model? | Eligibility, compatibility and pricing in the catalog are readable; the same rules in Apex are invisible to the agent | Move rules back to configuration — which is worth doing regardless of AI |
| Is there something for it to do? | An agent that can only explain moves work to a human rather than removing it | Design real actions with the boundary stated: raise the order, apply the credit, book the visit |
| Is the data trustworthy? | Asset records corrupted by cancel-and-rebuild workarounds make an agent confidently wrong at scale | Fix the asset integrity problem first; agents amplify data quality in both directions |
| Can you measure the result? | Without containment, resolution and escalation rates against a baseline, a pilot cannot be judged and will not scale | Establish the baseline before deployment, not after the first optimistic demo |
This is why AI readiness work so often turns out to be catalog and data work. That is not a reason to delay — it is work with its own return, whether or not you deploy a single agent, and it is the same work that makes the platform cheaper to change.
The work that makes an agent useful
Mostly data, permissions and measurement. The model is the part that already works.
Use-Case Selection & Readiness
Choosing where an agent will genuinely help, and being explicit about what has to be true first.
- Use-case assessment against the readiness test
- Value and volume sizing per candidate
- Data and access gap analysis
- Sequencing: what to fix before deploying
- An honest no when the case is not there
Data Grounding
Making the state an agent needs readable, current and trustworthy — the largest and least glamorous part of the work.
- Order, asset and contract state surfacing
- Billing and usage data access design
- Catalog rules moved into the readable model
- Retrieval design and relevance tuning
- Data quality remediation before deployment
Action Design
Defining what the agent actually does, so it completes work rather than describing it.
- Action inventory and boundary definition
- Wiring actions into catalog and order flows
- Confirmation and reversal behaviour
- Partial-completion and handoff design
- Reuse of existing journeys rather than new ones
Guardrails & Escalation
The limits, decided before deployment rather than after the first expensive outcome.
- Authority limits and approval thresholds
- Regulatory and commercial constraints
- Escalation triggers and human handoff
- Audit logging of every agent action
- Review process for boundary cases
Measurement & Tuning
Proving whether it worked, in numbers agreed before the pilot rather than chosen afterwards.
- Baseline capture before deployment
- Containment, resolution and escalation rates
- Quality review of agent outcomes
- Failure-mode analysis and retuning
- Scale-or-stop decision on evidence
Operating Model & Adoption
Fitting the agent into how teams already work, rather than creating a fourth version of the truth.
- Agent and human role definition
- Care and sales workflow integration
- Team enablement and expectation setting
- Ownership of agent behaviour over time
- Change process as the catalog evolves
Small, measured, and stopped early if it does not work
Agentic AI rewards a narrow first deployment with real measurement far more than a broad one with an impressive demo.
Readiness & Use-Case Selection
Candidate use cases against the readiness test, sized by volume and value. The output frequently includes a recommendation to fix grounding before deploying anything, which is the cheapest advice on this page.
Ground the Data
Surfacing order, asset, contract and billing state, moving catalog rules back into the readable model, and remediating the data quality problems that would otherwise be amplified. This is usually the longest step and it has value on its own.
Design Actions & Guardrails
What the agent may do, up to what limit, with what confirmation, and where it must stop and escalate. Agreed with the people accountable for the outcomes — care, commercial and compliance, not only architecture.
Deploy Narrow
One use case, one segment, with the baseline captured beforehand. Containment, resolution, escalation and outcome quality are tracked from day one, and the agent's actions are auditable.
Scale or Stop
On evidence. If the numbers hold, extend to adjacent use cases reusing the same grounding. If they do not, we say so and stop — a quiet pilot that nobody cancels is the most common outcome in this space and the most wasteful.
Grounding work overlaps with Communications Cloud optimization and customization, because rules in code are invisible to agents. Order state comes from order management & automation.
What a grounded agent actually changes
Described as operating changes. The numbers that matter are the ones measured on your own baseline, not ones published on a page.
Responses come from real state
Order, asset and contract records are grounded, so an agent answers from what the customer actually has rather than from a plausible generalisation.
Work is completed, not forwarded
Designed actions mean the agent raises the order, applies the credit or books the visit — so deflection reflects work removed rather than work moved.
One answer across channels
Agents read the same catalog, order and asset state that care and self-serve read, so a customer cannot get three different answers depending on where they asked.
Limits exist before incidents do
Authority thresholds, approval steps and escalation triggers are defined up front and every action is logged, which is what makes agentic AI viable in a billing-connected environment.
The pilot can actually be judged
Containment, resolution and escalation measured against a baseline captured beforehand, so scaling is a decision rather than an act of faith.
The groundwork pays either way
Readable catalog rules, trustworthy asset records and surfaced order state make the platform cheaper to change whether or not a single agent is deployed.
Related Twopir work
Related: Salesforce Communications Cloud consulting for the wider practice, and Salesforce for telecom infrastructure for operator systems more broadly. Named communications-provider references are discussed under NDA in a scoping call.
We will tell you when the answer is not an agent
A meaningful share of telecom AI use cases are better solved by fixing the catalog model or the order journey. Those fixes are cheaper, they work immediately, and they are also what any future agent would have needed anyway.
We test readiness before proposing a use case
Five questions about grounding, rules, actions, data quality and measurement. A no is not a reason to stop — it is the actual project, and it usually has its own return.
We treat grounding as the main work
Making order, asset, contract and billing state readable and trustworthy is most of the effort and nearly all of the risk. The model is the part that already works.
We design guardrails before deployment
Authority limits, approval thresholds and escalation paths agreed with care, commercial and compliance. In a billing-connected environment, discovering the limits afterwards is expensive.
We baseline before we deploy
Containment, resolution and escalation measured against a real starting point. A pilot without a baseline can be declared a success indefinitely and never scales.
We stop pilots that are not working
A quiet pilot nobody cancels is the most common and most wasteful outcome in agentic AI. We build the stop decision into the plan, with the evidence that would trigger it.
Answers before the first call
Agentforce for Communications is Salesforce's set of prebuilt, telecom-specific AI agents. They run on the Agentforce platform and connect to CRM, OSS and BSS systems so they can read live customer, service and billing data and act on it, rather than answering from a general model. Salesforce has announced five: Billing Resolution, which analyses bills across fragmented third-party data and helps resolve disputes within defined guardrails; Quoting, which generates complex telecom quotes from predefined business rules; Site Grouping, which organises multi-location deals from natural-language prompts; Service Level Objective Insights, which compares real-time network usage against compliance criteria; and Guided Selling, which helps field technicians identify upsell opportunities onsite. Availability differs by agent and changes, so confirm the current position against Salesforce's own product pages before committing.
The prebuilt Communications agents are designed around the telecom data model and processes that Communications Cloud provides — catalog, quoting, orders, assets and contracts. They are grounded in that state, which is what separates them from a general-purpose assistant. You can deploy general Agentforce capability on Salesforce without Communications Cloud, but the telecom-specific agents assume the telecom-specific model underneath them. If you are running standard CRM with a custom telecom data model, the honest answer is that it depends on how closely your model resembles the standard one, and that is worth assessing before licensing anything.
Because an agent can only act on what it can read, and in most telecom estates the things it needs to read are either spread across systems it has no grounded access to, or expressed in code rather than in the data model. Eligibility, compatibility and pricing rules held in a well-modelled catalog are readable by an agent; the same rules implemented in Apex are invisible to it. Asset records corrupted by cancel-and-rebuild workarounds make an agent confidently wrong at scale, because agents amplify data quality in both directions. Fixing those things is the project, and it is worth doing whether or not an agent is ever deployed — which is why we sequence it first rather than treating it as a blocker.
Guardrails designed before deployment rather than after an incident, which in practice means four things. Authority limits: the maximum credit, adjustment or commitment the agent can make unaided. Approval thresholds: what requires a human to confirm before it happens. Escalation triggers: the conditions under which the agent must stop and hand off, including uncertainty about its own grounding. And audit logging of every action, so behaviour can be reviewed and boundary cases fed back into the design. These are decided with the people accountable for the outcomes — care, commercial and compliance — not by architecture alone, because they are business risk decisions wearing a technical label.
Against a baseline captured before deployment, using containment rate, resolution rate, escalation rate and outcome quality. Conversation volume is the metric to avoid: it tells you the agent was used, not that anything improved. The distinction that matters most is between work removed and work moved — an agent that explains a billing dispute and hands it to a human has not reduced effort, it has added a step, and deflection measured carelessly will show that as a success. We agree the measures and capture the baseline before the pilot, precisely so the result cannot be reinterpreted afterwards.
Narrow, measured, and with a stated stop condition. One use case, one segment, with the baseline captured beforehand and containment, resolution and escalation tracked from day one. Broad pilots produce impressive demonstrations and ambiguous evidence, and they tend to persist indefinitely because nobody can prove they are not working. We build the scale-or-stop decision into the plan with the evidence that would trigger each, because a quiet pilot nobody cancels is the most common and most wasteful outcome in this area.
Salesforce builds and licenses Agentforce and the prebuilt Communications agents; the platform, the models and the agent framework are Salesforce capabilities. Twopir Consulting does the implementation work that determines whether they are useful in your org: assessing which use cases are ready, grounding the order, asset, contract and billing state the agents need, moving catalog rules into a form an agent can read, designing the actions and the guardrails, running a measured pilot against a real baseline, and making the scale-or-stop call on evidence. Twopir Consulting is a Salesforce Partner. We do not resell licences, and we will tell you when the answer to your problem is a better catalog model rather than an agent.
Next Step
Tell us the use case. We will tell you what has to be true first.
An AI readiness assessment tests candidate use cases against grounding, rules, actions, data quality and measurability — and gives you a sequenced plan, including the parts worth doing whether or not you deploy an agent.
Speak with a team that will tell you when it is not an AI problem
