July 2026
TL;DR
AI conversations today tend to split into two distinct groups. One is still working through whether to adopt AI at all. The other is several years further along, weighing token costs, agent reliability, and whether the return justifies the spend. Most founders only see the room they happen to be standing in, and most advisors only speak to one. Saguna Consulting sits in both, not as a matter of positioning, but because the same AI systems used with clients, for content, research, prospecting, and training, are also the systems the firm runs on internally.
)%20(20).png)
The headlines suggest AI adoption is nearly universal. The underlying data tells a more uneven story.
The OECD's 2026 D4SME survey of small and mid-size businesses found that 61 percent now use at least one AI application. Among that group, however, roughly three in four remain what the survey terms "AI novices," running a single tool for one isolated task. Fewer than 4 percent have integrated AI into how the business actually operates.
Larger organizations are not meaningfully further along. According to Mayfield's 2026 CXO AI Survey of 266 CIOs, CTOs, Chief AI Officers, and CISOs, close to eight in ten organizations report having adopted agentic AI, systems that plan and execute multi-step work with limited human input, somewhere in the business. Fewer than one in ten have it running in production. The distance between "we tried it" and "we depend on it" is where most companies currently sit.
That gap is not a technology failure. It is a sequencing failure: organizations acquire the tool before defining the workflow it should support, or issue a mandate to "use more AI" without first agreeing on what a good outcome looks like.
Four causes recur across engagements, regardless of industry or company size:
None of these are technology problems. They are operating-discipline problems, which places this squarely in the domain of consulting rather than procurement.
Saguna Consulting treats AI as infrastructure, threaded through how work gets done, rather than layered on top of an unchanged process. Internally, this is visible in how the firm builds its content pipelines, trains research analysts, manages outbound prospecting, and structures its own go-to-market systems. These practices were not the output of a formal digital-transformation initiative; they were built into daily operations directly, in the same way a client would be expected to build them. That operating experience is what informs the firm's approach with a founder moving from idea to MVP, or an enterprise team deciding what to automate first.
For the past two years, the working assumption has been that AI gets cheaper every year, and that bills should shrink accordingly. That assumption does not hold in production settings. Per-token pricing has fallen, but the cost of running AI in production has risen, because agentic systems plan, retry, call external tools, verify their own output, and resend the full conversation history at each step. Cost does not scale linearly with usage; it compounds with the complexity of the task.
The more expensive part of an AI-driven task is typically not the first response, but everything that follows: verification, re-running, and the human judgment required to decide whether the output can be trusted. Two runs of an identical task can produce meaningfully different costs, since an autonomous system rarely follows the same execution path twice. Budgeting for AI as though it were a fixed-price tool tends to produce inaccurate forecasts.
Two opposing responses to this cost pressure have both proven flawed. One approach treated heavy usage as a proxy for value, in some organizations, to the point of ranking employees by token consumption, which incentivized higher usage rather than better output. The opposing response has been to cut usage sharply, shrinking context windows and restricting access to more capable models. Some of that discipline is warranted, but past a certain point, the cuts remove the context that made the output usable in the first place. The visible bill falls, and the underlying cost resurfaces as rework and escalation.
Token volume is a measure of activity, not of value delivered. The more useful question is whether the work cleared the bar required of it. A workflow that costs more per run but rarely requires a second pass is, in most cases that matter, less expensive than one that appears efficient on paper but is discarded half the time.
The Model Context Protocol (MCP), the open standard connecting AI agents to external tools, became infrastructure over the course of 2026, with thousands of servers now in use across major platforms. That growth carries a cost of its own: a system that reads hundreds of tool definitions on every call consumes budget before it has completed any work.
Looking further out, Gartner estimates that by 2030, roughly a fifth of enterprise software spending will be exposed to what it terms "agentic arbitrage," AI agents completing tasks directly and removing the interfaces that previously sat between the user and the outcome. The revenue model underlying the software industry is being revised as a result.
None of this argues for adopting AI immediately and everywhere, or for withholding it altogether. The organizations still standing several years from now are unlikely to be the ones that used the most AI, or the least. They will be the ones able to answer a single question with confidence: did this create value, and can that be proven? That is the starting point on every Saguna Consulting engagement, and where the conversation would begin with any organization asking the same question of itself.
The headlines suggest AI adoption is nearly universal. The underlying data tells a more uneven story.
The OECD's 2026 D4SME survey of small and mid-size businesses found that 61 percent now use at least one AI application. Among that group, however, roughly three in four remain what the survey terms "AI novices," running a single tool for one isolated task. Fewer than 4 percent have integrated AI into how the business actually operates.
Larger organizations are not meaningfully further along. According to Mayfield's 2026 CXO AI Survey of 266 CIOs, CTOs, Chief AI Officers, and CISOs, close to eight in ten organizations report having adopted agentic AI, systems that plan and execute multi-step work with limited human input, somewhere in the business. Fewer than one in ten have it running in production. The distance between "we tried it" and "we depend on it" is where most companies currently sit.
That gap is not a technology failure. It is a sequencing failure: organizations acquire the tool before defining the workflow it should support, or issue a mandate to "use more AI" without first agreeing on what a good outcome looks like.
Four causes recur across engagements, regardless of industry or company size:
None of these are technology problems. They are operating-discipline problems, which places this squarely in the domain of consulting rather than procurement.
Saguna Consulting treats AI as infrastructure, threaded through how work gets done, rather than layered on top of an unchanged process. Internally, this is visible in how the firm builds its content pipelines, trains research analysts, manages outbound prospecting, and structures its own go-to-market systems. These practices were not the output of a formal digital-transformation initiative; they were built into daily operations directly, in the same way a client would be expected to build them. That operating experience is what informs the firm's approach with a founder moving from idea to MVP, or an enterprise team deciding what to automate first.
For the past two years, the working assumption has been that AI gets cheaper every year, and that bills should shrink accordingly. That assumption does not hold in production settings. Per-token pricing has fallen, but the cost of running AI in production has risen, because agentic systems plan, retry, call external tools, verify their own output, and resend the full conversation history at each step. Cost does not scale linearly with usage; it compounds with the complexity of the task.
The more expensive part of an AI-driven task is typically not the first response, but everything that follows: verification, re-running, and the human judgment required to decide whether the output can be trusted. Two runs of an identical task can produce meaningfully different costs, since an autonomous system rarely follows the same execution path twice. Budgeting for AI as though it were a fixed-price tool tends to produce inaccurate forecasts.
Two opposing responses to this cost pressure have both proven flawed. One approach treated heavy usage as a proxy for value, in some organizations, to the point of ranking employees by token consumption, which incentivized higher usage rather than better output. The opposing response has been to cut usage sharply, shrinking context windows and restricting access to more capable models. Some of that discipline is warranted, but past a certain point, the cuts remove the context that made the output usable in the first place. The visible bill falls, and the underlying cost resurfaces as rework and escalation.
Token volume is a measure of activity, not of value delivered. The more useful question is whether the work cleared the bar required of it. A workflow that costs more per run but rarely requires a second pass is, in most cases that matter, less expensive than one that appears efficient on paper but is discarded half the time.
The Model Context Protocol (MCP), the open standard connecting AI agents to external tools, became infrastructure over the course of 2026, with thousands of servers now in use across major platforms. That growth carries a cost of its own: a system that reads hundreds of tool definitions on every call consumes budget before it has completed any work.
Looking further out, Gartner estimates that by 2030, roughly a fifth of enterprise software spending will be exposed to what it terms "agentic arbitrage," AI agents completing tasks directly and removing the interfaces that previously sat between the user and the outcome. The revenue model underlying the software industry is being revised as a result.
None of this argues for adopting AI immediately and everywhere, or for withholding it altogether. The organizations still standing several years from now are unlikely to be the ones that used the most AI, or the least. They will be the ones able to answer a single question with confidence: did this create value, and can that be proven? That is the starting point on every Saguna Consulting engagement, and where the conversation would begin with any organization asking the same question of itself.
The real challenge isn't adopting AI; it's making it work consistently, securely, and at scale. Let's build an AI strategy that delivers value long after the pilot ends.
.png)
