CRM trends:
data and actions
We examine AI assistants, shared customer history and automated actions. Check data, limits and results before scaling up.

Check the actions.
Measure the result.
Practical workflows, official guides and pilot criteria for developing CRM without unsupported promises.
What is changing in CRM and how to choose a useful direction
In 2026, CRM is increasingly seen as a place to carry out work: it helps interpret enquiries, prepare the next step and connect actions across departments. Buyers need to distinguish an available feature from an implementation result. AI included in a subscription does not mean that salespeople receive accurate suggestions or stop losing deals.
This guide was prepared on September 24, 2026. Product capabilities were checked against the documentation linked below. Plans, availability and control names may change. The teaching scenarios demonstrate how to verify results; they are not iSEO.pro client case studies.
| Direction | When it helps | What to check |
|---|---|---|
| AI for conversations and records. | Salespeople spend too long recording agreements. | Field accuracy and time spent checking the result. |
| Automated actions. | Enquiries wait to be assigned or moved to the next step. | Trigger conditions, exceptions and duplicate prevention. |
| Unified customer data. | History is fragmented across channels and systems. | Correct matching and employee access. |
| Analytics and retention. | Decisions rely on an incomplete picture. | Metric definitions, sources and forecast usefulness. |
Start with an observable problem. If enquiries never reach CRM, restore the transfer first. AI lead scoring on an incomplete flow describes only the demand the system has seen.
A unified customer record needs matching rules
One customer can have several contacts, legal entities and communication channels. Merging every record with a matching phone number is risky: a shared company number does not always identify one person. Before automating, define what your business considers a customer, a contact and a separate deal.
For example, the Bitrix24 duplicate-management documentation describes different conditions for automatic and manual merging, including restrictions involving owners and stages. The feature is not available on every plan. This illustrates why one CRM’s rules cannot simply be applied to every project.
- Prepare test pairs: one customer in two channels, two employees of one company, a shared phone number and a changed email address.
- In a test environment, open the contacts section and your CRM’s duplicate-control settings.
- Review the suggested matches and the values that will remain after merging.
- Check linked deals, communication history and the responsible employee.
- Agree how to handle ambiguous cases before bulk processing.
Assign owners for reference data and required fields. Quality control should identify specific errors, not just count completed records: outdated status, a missing next step or an incorrect company link. Data preparation is covered in the CRM implementation plan.
Check how agreements are extracted, not how polished the summary sounds
Transcription, summarisation and field completion solve different tasks. A well-written summary can conceal an incorrect quantity, date or address. These values matter to sales because they flow into the proposal and subsequent actions.
In the Bitrix24 CoPilot guide for CRM, call processing starts from the record’s timeline. The service produces a transcript, summary and field data. Replacing existing values requires confirmation. Before a pilot, check the feature’s conditions, employee access and suitable recordings.
The customer changes the equipment quantity
- Starting data
- The call starts with four pumps. Later, the customer clarifies that six are needed and the delivery date is still to be agreed.
- Correct result
- The quantity is six. No delivery date is invented. The next task records the need to agree it.
- Check
- Compare the fields with the recording and check that the old value has not remained in a linked document.
Include noisy recordings, interruptions, negation, changed decisions and unknown specifications in the test set. Score each important field separately. If the salesperson has to replay the whole call, the actual time saved may be smaller than expected.
Separate suggestions, preparation and execution
The next level of automation is taking action in CRM: creating a task, changing the owner, completing a field or moving a deal. These capabilities are already described, for example, in the Kommo AI agent guide. Their availability does not remove the need to configure trigger conditions and check the consequences.
| Mode | Example | What to agree |
|---|---|---|
| Suggestion. | Suggest the next step to a salesperson. | The basis for the suggestion and an option to reject it. |
| Preparation. | Draft a reply or task. | Who reviews the content before it is applied. |
| Rule-based execution. | Assign an owner by region and queue. | Allowed values, exceptions and an action log. |
| Changes to important terms. | Change an agreed discount or order status. | Authority, confirmation and a way to correct the result. |
A language model is not always needed for a simple rule. Standard automation can assign enquiries using a known region field. AI helps interpret free text, but the system still needs to enforce business constraints.
Ask for the same event to be repeated twice and simulate a delayed integration response. The agreed number of tasks and changes should result. Duplicate operations are not an acceptable price for a “smart” assistant.
Omnichannel means a continuous history, not a collection of icons
Connected email, phone and messaging help when employees understand the customer’s history and can continue the work. Test the full journey: website enquiry, salesperson’s reply, clarification in another channel, order change and final record.
Agree how the link between a messaging user and a CRM record is verified. Do not reveal order details merely because someone gives a similar name or number. Different actions may require different verification methods.
When work passes between employees, preserve the reason for contact, decisions already made and the next step. Define the queue, working hours and what happens when no salesperson is free. Saying “we’ll connect you now” without creating a request and assigning an owner is not a handoff.
Before implementation, check channel availability for your account, provider limits and charges. You do not need every messaging service at once: first make the workflow reliable in the channels your customers use.
Forecasting needs reliable history and comparison with actual results
AI can help explain a report or estimate the likelihood of an outcome, but the underlying definitions must be agreed. What counts as a qualified enquiry, paid deal, loss or repeat purchase? Changing these rules makes comparisons between periods ambiguous.
Check historical data before forecasting. Backdated deal closures, missing loss reasons and unrecorded enquiries distort the picture. If a model uses information that only appears after the outcome, testing can look better than real operation.
- Choose a completed period that was not used to tune the forecast.
- Compare the forecast with actual results and a simple, understandable baseline.
- Break down errors by segment rather than relying only on an overall average.
- Check what decision the employee makes based on the forecast.
A natural-language report should lead back to its source data and calculation period. If it says “sales fell”, the manager must be able to inspect the deals and calculation rules behind that conclusion. Connections to advertising sources are considered alongside business analytics.
Personalisation starts with a relevant next action
Retention is more than sending an automated message to everyone who has not bought recently. Retail and industrial equipment sales have different buying cycles. A seven-day return metric may help one product and reveal almost nothing about another.
Define an event after which a message actually helps: completed training, an agreed service date approaching or a needed item becoming available. Account for open requests and existing agreements. Do not propose a new purchase while the customer is waiting for a problem with the previous one to be resolved.
Check frequency limits, channel preferences and stopping messages after an opt-out. Compare similar groups to measure impact, including the final action, complaints and support costs. Sending more messages alone does not demonstrate value.
Test AI segmentation on understandable examples first. A salesperson should know why a customer is in a segment and be able to correct the source record. A data error must not silently spread to every campaign.
Low-code speeds up configuration but does not remove operational work
Builders help assemble a process from ready-made actions. Complexity appears at system boundaries: one record updates, another system rejects the request, or an employee edits a field manually. Each connection therefore needs an owner and a clear recovery procedure.
Record which system owns the price, payment status, availability and responsible employee. Define the acceptable synchronisation delay. Test repeated events, unavailable services and conflicting changes. Errors must be visible to an employee who knows how to handle them.
Useful integration logs include the time, operation ID, outcome and rejection reason. Do not copy unnecessary conversation content or access secrets into them. Once a fault is fixed, a missed event should be recoverable without creating the same entity again.
Rules must not exist only in one integrator’s head. Prepare a connection diagram, field descriptions and support instructions. Common failures are covered in the guide to CRM implementation problems.
Test mobile access and roles against real work
A field salesperson may need a short workflow: find a customer, review an agreement, record the meeting outcome and set the next task. A full copy of a complex desktop record is not always helpful. Establish which actions are actually performed away from the office.
Check permissions using an ordinary employee role. Administrators see more, so a successful demonstration from their account does not prove the team’s workflow works. Test access to other employees’ deals, search, export and connected assistants within the agreed scope.
If operation with an unstable connection is promised, demonstrate a specific workflow: enter a note, lose connectivity, reconnect and synchronise. Clarify what data is stored locally and how conflicts are handled. Do not promise that the entire product works offline without testing it.
Train people on their role’s tasks, including failures and corrections. Success means that employees complete the action independently and know where to get help when the system behaves differently.
How to run a CRM pilot with verifiable results
Choose one process and team whose current work can be measured. Record enquiry volumes, processing time, errors and manual corrections. Then define pilot criteria and stopping rules if the new feature makes the result worse.
- Prepare test data and difficult cases.
- Test the feature without affecting the main workflow.
- Train participants and assign an owner for the result.
- Launch at a limited volume and log errors.
- Compare quality, time and costs with the original process.
Include licences, connected services, AI processing, integrations, answer review and support in the calculation. Estimate data cleaning separately: it can be the main driver of effort. See the CRM budget breakdown.
Scale a workflow that has demonstrated value. If the pilot does not help, identify why: the wrong task, insufficient data, product limits or awkward employee workflows. Decisions about further CRM development should be based on those observations.
Is your CRM ready for the new workflow?
Frequently asked questions about CRM trends
Do you need to change CRM just to get AI?
First check what your current product and a separate integration can do. Migration should address verified limitations; its cost includes data, processes, training and the transition period.
Can AI complete every record without review?
Some fields can be automated, but suitability depends on accuracy and the consequences of mistakes. Test important values on difficult examples and require confirmation where needed.
Which matters more: the agent or data quality?
If records are incomplete, contradictory or incorrectly linked, fix that foundation first. An agent can speed up work with good data but cannot guarantee recovery of unknown facts.
Does an omnichannel CRM merge all contacts automatically?
Not always. You need identification rules, channel configuration and a process for ambiguous matches. Check that the history belongs to the right person and is accessible only to appropriate roles.
How should a sales forecast be evaluated?
Compare it with actual results over an independent period and with a clear baseline calculation. Break down errors by segment and check how the forecast affects employee decisions.
Where should a small business start?
Start with a recurring problem: lost enquiries, no next step or manual data transfer. Agree a limited workflow and measure its result before extending automation.
LET’S PLAN THE NEXT STAGE
OF YOUR CRM
We examine processes, data quality and system limits, then define a priority workflow, pilot scope and acceptance criteria.