<p><strong>Executive Summary</strong></p><ul><li><p><strong>The gap between</strong> AI proof-of-concept and embedded capability is a data readiness and change management problem.</p></li><li><p><strong>Finance teams</strong> that upskilled before AI matured are outperforming those that waited; capability-building must precede, not follow, tool deployment.</p></li><li><p><strong>AI in finance</strong> is most effective when deployed against high-volume, rules-bound processes, where accuracy gains are measurable.</p></li><li><p><strong>Prompt engineering</strong> is a key discipline: standardising prompts produces more consistent, business-specific outputs.</p></li><li><p><strong>Human-in-the-loop design</strong> is the governing principle for AI in financial processes: human judgement is reserved for exceptions, validation and contextual decisions.</p></li><li><p><strong>Finance’s long-term value</strong> lies not in the volume of analysis it produces but in its ability to interpret data as business narrative, a skill AI augments but cannot replace.</p></li></ul>.<p>Finance leaders no longer question whether artificial intelligence (AI) belongs inside the function. Instead, they ask why, given the sheer scale of investment in various tools and pilots, genuine transformation remains so rare. Across industries, finance functions are awash in proofs-of-concept; what they lack is the organisational architecture to make them stick. Against this backdrop, Raghuram Krishnan, Senior Director – Finance Operations & Analytics at Adobe, examined what a sustained AI journey inside a complex global finance function actually looks like, where the real obstacles lie and what other organisations can draw from the experience.</p><h2><strong>The Starting Condition: Data, Not Algorithms</strong></h2><p>Often, a key first-level obstacle is not an absence of <em>capable tools</em> but an absence of <em>useful data</em>. Until quite recently, <strong>Adobe's</strong> finance function operated across fragmented systems that did not communicate well, with unclear distinctions between confidential and restricted information, and with a collective assumption (common across finance teams) that its data quality was broadly adequate. The discipline of mapping what data existed, how clean it was, what could be used, and what needed masking, consumed significant early effort but remains ongoing. The lesson is clear: AI surfaces data problems rather than hiding them, and organisations that have not done foundational data work will find that every pilot hits the same ceiling.</p><p>A parallel exercise, cataloguing every IT-approved software tool across the organisation, revealed that Adobe's finance function had access to 72 distinct tools, of which most team members knew only 2 or 3. The inventory exercise became a primary source of capability. Knowing what existed, and securing guidance from IT and security teams on what was cleared for use, meant that the team could make deliberate choices rather than default to the most visible tools in the market.</p><h2> <strong>Execution at Scale: The Use Cases That Delivered</strong></h2><p>Several projects at Adobe shared a common architecture: high-volume, rules-bound processes where manual handling created bottlenecks, followed by AI deployment with human oversight at decision points. Predictive revenue modelling, built on SQL and connected to the CRM pipeline using Python written entirely through Copilot and ChatGPT, now operates at 97-99% accuracy on a weekly basis. Such levels of accuracy did not arrive immediately, but were built over 18 months as business model changes were absorbed and the model was refined. The implication is that accuracy in AI-driven forecasting is an outcome of iteration, not of implementation alone.</p><p>A chatbot built for quarterly business reviews replaced a 100-slide deck with 4 summary slides and a conversational interface. The CEO can now interrogate which markets grew, which product lines underperformed and where exceptions lay, in real time, without sequencing through a fixed presentation. Adobe ran its first version of this system 3 years ago, which places it considerably ahead of the median on the deployment curve. Its new accounts-receivable summary, rebuilt inside Claude in 5 days after years of manual Excel assembly, demonstrates how far the tooling has moved: what was projected as a year-long project was prototyped within a week.</p><p>A successful deal-reconciliation project has addressed one of finance's most time-sensitive processes – validating 40,000 annual deals for IRR calculation and commission payments. Previous automation through bots failed because 60-70% of deals tend to be non-standard. An LLM trained on Adobe's booking policy can now handle the nuances that rule-based automation could not. The design principle holds: standard deals are handled automatically; non-standard deals involve human review. As the system learns, the proportion requiring human intervention is expected to fall.</p><h2><strong>The Upskilling Imperative</strong></h2><p>One key difference separates Adobe's trajectory from the common experience of stalled AI adoption: the team was upskilled <em>before</em> the AI wave arrived, not <em>in response to it</em>. From 2022, the finance function set individual learning goals; by 2023, 65% of the team was pursuing courses – including Python, SQL and other skills not conventionally associated with finance. The result was a cohort of analysts who began solving their own problems autonomously, writing code through Copilot, modifying outputs and building local efficiencies. By 2024 and 2025, the function absorbed the equivalent of ten roles' worth of work without adding headcount. The 2026 target is 90% of the team enrolled in active courses.</p><p>The mechanics of sustaining this culture are deliberate: learning progress is displayed at every all-hands meeting, leaders submit their own course certificates to their teams and failures are shared as case studies rather than suppressed. A hackathon run within the finance function alone identified savings of 10,000 hours. Adobe recognised that capability-building is a management discipline with measurable outcomes, not a cultural aspiration.</p><h2><strong>Prompt Engineering as Organisational Infrastructure</strong></h2><p>An often-underappreciated fact is that AI tools give generic outputs in response to generic inputs. For a finance function managing 70% of a company's operating expenditure and 95% of its headcount, consistency of analysis across a large team is a governance requirement. Adobe has responded by treating prompt engineering as a vertical within the organisation, with a dedicated effort to build, test and standardise prompts by business type: one prompt set for go-to-market analysis, another for engineering cost centres, another for top-line review.</p><p>The practical effect is that any analyst drawing on the prompt library produces analysis with comparable structure, framing and reliability. Cultural and linguistic variations across a global team, which can be a genuine source of inconsistency in analyst output, are managed through the prompt layer rather than individual supervision. This is a governance solution as much as a technology one.</p><h2><strong>The Finance Function's Strategic Horizon</strong></h2><p>Does AI make the finance function more strategic or more commoditised? Mr Krishnan's view is that it does both, in sequence. Efficiency gains reduce the function's transactional footprint; what fills the space is determined by how deliberately the function has built its adjacent capabilities. Adobe has run workshops specifically on translating numerical analysis into business narrative on the premise that data is now abundant but interpretation remains scarce. The analyst who can explain why a number moved, in language the business understands, is performing a function that AI currently augments rather than replaces.</p><p>In the longer term, as finance teams become leaner, their members will move outward – to functions that are earlier in the AI adoption curve, to organisations in other sectors that are beginning the journey, to roles where the combination of financial rigour and AI fluency is rare enough to command a premium. Mr Krishnan sees this not as a risk to the profession but as a multiplication of its reach. The teams that build this capability now are, in effect, training the next generation of finance leaders for an AI-native operating environment.</p>
<p><strong>Executive Summary</strong></p><ul><li><p><strong>The gap between</strong> AI proof-of-concept and embedded capability is a data readiness and change management problem.</p></li><li><p><strong>Finance teams</strong> that upskilled before AI matured are outperforming those that waited; capability-building must precede, not follow, tool deployment.</p></li><li><p><strong>AI in finance</strong> is most effective when deployed against high-volume, rules-bound processes, where accuracy gains are measurable.</p></li><li><p><strong>Prompt engineering</strong> is a key discipline: standardising prompts produces more consistent, business-specific outputs.</p></li><li><p><strong>Human-in-the-loop design</strong> is the governing principle for AI in financial processes: human judgement is reserved for exceptions, validation and contextual decisions.</p></li><li><p><strong>Finance’s long-term value</strong> lies not in the volume of analysis it produces but in its ability to interpret data as business narrative, a skill AI augments but cannot replace.</p></li></ul>.<p>Finance leaders no longer question whether artificial intelligence (AI) belongs inside the function. Instead, they ask why, given the sheer scale of investment in various tools and pilots, genuine transformation remains so rare. Across industries, finance functions are awash in proofs-of-concept; what they lack is the organisational architecture to make them stick. Against this backdrop, Raghuram Krishnan, Senior Director – Finance Operations & Analytics at Adobe, examined what a sustained AI journey inside a complex global finance function actually looks like, where the real obstacles lie and what other organisations can draw from the experience.</p><h2><strong>The Starting Condition: Data, Not Algorithms</strong></h2><p>Often, a key first-level obstacle is not an absence of <em>capable tools</em> but an absence of <em>useful data</em>. Until quite recently, <strong>Adobe's</strong> finance function operated across fragmented systems that did not communicate well, with unclear distinctions between confidential and restricted information, and with a collective assumption (common across finance teams) that its data quality was broadly adequate. The discipline of mapping what data existed, how clean it was, what could be used, and what needed masking, consumed significant early effort but remains ongoing. The lesson is clear: AI surfaces data problems rather than hiding them, and organisations that have not done foundational data work will find that every pilot hits the same ceiling.</p><p>A parallel exercise, cataloguing every IT-approved software tool across the organisation, revealed that Adobe's finance function had access to 72 distinct tools, of which most team members knew only 2 or 3. The inventory exercise became a primary source of capability. Knowing what existed, and securing guidance from IT and security teams on what was cleared for use, meant that the team could make deliberate choices rather than default to the most visible tools in the market.</p><h2> <strong>Execution at Scale: The Use Cases That Delivered</strong></h2><p>Several projects at Adobe shared a common architecture: high-volume, rules-bound processes where manual handling created bottlenecks, followed by AI deployment with human oversight at decision points. Predictive revenue modelling, built on SQL and connected to the CRM pipeline using Python written entirely through Copilot and ChatGPT, now operates at 97-99% accuracy on a weekly basis. Such levels of accuracy did not arrive immediately, but were built over 18 months as business model changes were absorbed and the model was refined. The implication is that accuracy in AI-driven forecasting is an outcome of iteration, not of implementation alone.</p><p>A chatbot built for quarterly business reviews replaced a 100-slide deck with 4 summary slides and a conversational interface. The CEO can now interrogate which markets grew, which product lines underperformed and where exceptions lay, in real time, without sequencing through a fixed presentation. Adobe ran its first version of this system 3 years ago, which places it considerably ahead of the median on the deployment curve. Its new accounts-receivable summary, rebuilt inside Claude in 5 days after years of manual Excel assembly, demonstrates how far the tooling has moved: what was projected as a year-long project was prototyped within a week.</p><p>A successful deal-reconciliation project has addressed one of finance's most time-sensitive processes – validating 40,000 annual deals for IRR calculation and commission payments. Previous automation through bots failed because 60-70% of deals tend to be non-standard. An LLM trained on Adobe's booking policy can now handle the nuances that rule-based automation could not. The design principle holds: standard deals are handled automatically; non-standard deals involve human review. As the system learns, the proportion requiring human intervention is expected to fall.</p><h2><strong>The Upskilling Imperative</strong></h2><p>One key difference separates Adobe's trajectory from the common experience of stalled AI adoption: the team was upskilled <em>before</em> the AI wave arrived, not <em>in response to it</em>. From 2022, the finance function set individual learning goals; by 2023, 65% of the team was pursuing courses – including Python, SQL and other skills not conventionally associated with finance. The result was a cohort of analysts who began solving their own problems autonomously, writing code through Copilot, modifying outputs and building local efficiencies. By 2024 and 2025, the function absorbed the equivalent of ten roles' worth of work without adding headcount. The 2026 target is 90% of the team enrolled in active courses.</p><p>The mechanics of sustaining this culture are deliberate: learning progress is displayed at every all-hands meeting, leaders submit their own course certificates to their teams and failures are shared as case studies rather than suppressed. A hackathon run within the finance function alone identified savings of 10,000 hours. Adobe recognised that capability-building is a management discipline with measurable outcomes, not a cultural aspiration.</p><h2><strong>Prompt Engineering as Organisational Infrastructure</strong></h2><p>An often-underappreciated fact is that AI tools give generic outputs in response to generic inputs. For a finance function managing 70% of a company's operating expenditure and 95% of its headcount, consistency of analysis across a large team is a governance requirement. Adobe has responded by treating prompt engineering as a vertical within the organisation, with a dedicated effort to build, test and standardise prompts by business type: one prompt set for go-to-market analysis, another for engineering cost centres, another for top-line review.</p><p>The practical effect is that any analyst drawing on the prompt library produces analysis with comparable structure, framing and reliability. Cultural and linguistic variations across a global team, which can be a genuine source of inconsistency in analyst output, are managed through the prompt layer rather than individual supervision. This is a governance solution as much as a technology one.</p><h2><strong>The Finance Function's Strategic Horizon</strong></h2><p>Does AI make the finance function more strategic or more commoditised? Mr Krishnan's view is that it does both, in sequence. Efficiency gains reduce the function's transactional footprint; what fills the space is determined by how deliberately the function has built its adjacent capabilities. Adobe has run workshops specifically on translating numerical analysis into business narrative on the premise that data is now abundant but interpretation remains scarce. The analyst who can explain why a number moved, in language the business understands, is performing a function that AI currently augments rather than replaces.</p><p>In the longer term, as finance teams become leaner, their members will move outward – to functions that are earlier in the AI adoption curve, to organisations in other sectors that are beginning the journey, to roles where the combination of financial rigour and AI fluency is rare enough to command a premium. Mr Krishnan sees this not as a risk to the profession but as a multiplication of its reach. The teams that build this capability now are, in effect, training the next generation of finance leaders for an AI-native operating environment.</p>
