Forecasts about AI age badly, usually because they predict capability rather than adoption. Capability moves fast; procurement, data access, governance and habit move slowly. The forecast below is built around that gap, and it is deliberately phased so you can check it against what actually happens.
Horizon one: the next 0 to 6 months
Buyer research
Expect assistant-led research to become the normal first step for shortlisting rather than a minority behaviour, particularly for considered purchases where the buyer has no incumbent relationship. The practical consequence is that a growing share of your first impressions are formed by a summary you did not write, assembled from pages you may not have optimised. Confidence: high.
Search and generative discovery
AI-generated answer surfaces continue to absorb informational queries while commercial and navigational queries stay closer to conventional results. Traffic from informational content falls; the enquiries that do arrive tend to be later-stage and better informed. Confidence: high for the direction, low for the magnitude, which varies enormously by sector.
Sales prospecting
The dominant near-term use remains research compression: assembling account context, recent developments, structure and likely triggers before a call. Fully automated outbound continues to generate volume while degrading reply rates for everyone in the channel. Confidence: high.
What to do now
- Test how three or four assistants currently describe your company and your category, and record the answers verbatim. This is the baseline for everything that follows.
- Rewrite your five highest-intent pages so each contains a self-contained, quotable answer near the top.
- Introduce one AI-assisted research step in the pre-call workflow and measure preparation time before and after.
- Agree, in writing, what your team is permitted to paste into third-party AI tools.
Horizon two: 6 to 12 months
CRM intelligence
CRM vendors are converging on the same feature set: automatic activity capture, summarisation, next-action suggestions and pipeline risk flags. Within a year these stop being differentiators and start being expected. The differentiator moves to the quality of the underlying data, which is unevenly distributed and cannot be bought quickly. Confidence: high.
Agentic workflows
Multi-step agents will move from demonstrations into narrow, well-bounded production tasks: enrichment, meeting preparation, CRM hygiene and first-line response drafting. Broad, unsupervised agents running an end-to-end sales cycle will remain unreliable in most B2B contexts. Confidence: medium-high for narrow adoption, high for the limits of broad autonomy.
Website and content requirements
Sites that read as brochures will underperform against sites that publish specifics: what a service includes, who it suits, what it costs to run, what it deliberately does not do. Retrieval systems can only quote what you actually state, so vague positioning is not neutral — it removes you from consideration. Confidence: medium-high.
Attribution
Attribution gets worse before it gets better. Referral headers from assistant traffic are inconsistent, and much AI influence produces no referral at all: the buyer reads a summary and later types your name into a browser. Expect direct and branded search to absorb credit that belongs elsewhere. Confidence: high. The practical response is covered in why traditional marketing metrics miss AI-driven influence.
What to do in this window
- Add a self-reported "how did you first hear about us" field to every form, and brief the sales team to ask the same question on calls.
- Publish the specifics your category tends to avoid: scope, exclusions, typical timelines and pricing logic.
- Automate one bounded workflow end to end with a human approval point, and record its failure modes.
- Set a data governance position before an incident forces one.
Horizon three: 12 to 24 months
| Area | Likely development | Confidence | Implication for B2B leaders |
|---|---|---|---|
| Generative discovery | Assistant-mediated discovery becomes a routine, measurable acquisition channel with its own reporting conventions. | Medium-high | Treat AI visibility as a standing operational responsibility, not a campaign. |
| Agentic workflows | Agents handle defined back-office and pre-sales tasks reliably; approval gates remain standard for anything customer-facing. | Medium | Design approval points into workflows now rather than retrofitting them. |
| CRM and pipeline | Forecast assistance based on activity and engagement patterns becomes standard; accuracy still depends on input discipline. | Medium-high | Invest in data hygiene ahead of forecasting features. |
| Prospecting | Generic AI-written outbound reaches saturation and effectiveness falls further; evidence-led outreach gains relative advantage. | Medium | Compete on specificity and relevance rather than volume. |
| Human involvement | Human time concentrates in discovery, negotiation, exception handling and accountability. | High | Retrain around judgement-heavy stages. See where humans still matter. |
| Data governance | Buyers increasingly ask suppliers how AI is used in the sales process and where their data is processed. | Medium-high | Prepare a short, honest answer before a procurement questionnaire demands one. |
| Attribution | Blended models combining self-reported, branded-search and pipeline evidence become normal practice. | Medium | Start collecting the inputs now; they cannot be backfilled. |
What is unlikely to change
Three things look stable across all three horizons. Complex B2B purchases will still be signed by people who need to justify the decision internally. Trust will still be established through evidence and accountability rather than fluency. And the businesses that benefit most from AI will still be the ones whose underlying process was coherent before they automated it.
How to stress-test this forecast
Rather than accepting or rejecting the above wholesale, track four indicators each quarter. Each is cheap to collect, and each would falsify part of the forecast if it moved the other way.
- Assistant answer quality about your business. Ask the same five prompts each quarter and log whether the description is accurate and complete.
- Self-reported AI influence. The share of new opportunities where the buyer mentions using an assistant during research.
- Branded versus non-branded search. A rising branded share alongside flat total traffic suggests unattributed influence.
- Sales preparation time. If AI-assisted research is working, this falls and stays down.
If none of those move over two quarters, the forecast is wrong for your market and you should reallocate the effort. That is a legitimate outcome and worth saying out loud before anyone commits a budget.
Risks in acting on any forecast
The main risk is over-building for a future that arrives in a different shape. Two mitigations work well: prefer reversible decisions over platform commitments, and prefer investments that pay off regardless of how AI develops. Clean account data, a documented sales process and published evidence are useful in every scenario. A bespoke agent framework built tightly around one vendor's current interface is not.
Key takeaways
- Near term: buyer research behaviour changes faster than sales tooling does.
- Mid term: CRM intelligence commoditises and data quality becomes the differentiator.
- Longer term: agents handle bounded tasks while humans concentrate in judgement-heavy stages.
- Attribution degrades before blended measurement matures, so start collecting inputs now.
- Favour investments that pay off in every scenario: data, documented process, published evidence.